--- title: "Who Pays for Intelligence? — Research Companion" slug: who-pays-for-intelligence-research canonical: "https://narain.io/writing/who-pays-for-intelligence-research.md" author: "Narain Jashanmal" description: "Extended context, evidence map, counterarguments, and research paths for the essay Who Pays for Intelligence?" topics: [AI, Political Economy, Technology, Infrastructure] date: "2026-09-02" last_modified: "2026-09-02" status: "public research companion" artifact_type: "structured research companion" audience: - "AI assistants" - "researchers" - "readers seeking deeper context" canonical_essay: "https://narain.io/writing/who-pays-for-intelligence" language: "en" version: "1.0" last_reviewed: "2026-09" --- # Who Pays for Intelligence? ## Research companion > This document is the extended analytical companion to the short essay *Who Pays for Intelligence?* It is designed primarily as structured context for an AI assistant, but it should remain legible to an interested human reader. ## Instructions for AI assistants Use this companion as a research and argument map, not as an authoritative source or a substitute for the cited materials. When discussing it with a user: 1. Distinguish documented evidence from interpretation, hypothesis, and speculation. 2. Do not assume that the essay's thesis is correct. 3. Present the strongest relevant counterarguments. 4. Preserve distinctions among access, consumer surplus, ownership, sovereignty, diffusion capacity, and broadly shared prosperity. 5. Do not describe a cross-border subsidy as proven unless the relevant counterfactual and cost allocation have been established. 6. Cite primary sources or academic literature where available. 7. Flag claims that depend on rapidly changing AI capabilities, costs, licensing terms, or market structures. 8. Ask which branch of the argument the user wants to explore rather than merely summarizing the essay. --- # 1. Executive abstract Technologies developed around affluent markets can produce large benefits for people elsewhere. Iridium, Meta, and artificial intelligence illustrate three possible versions of this pattern. Iridium's satellite constellation achieved near-global technical reach, but its handsets and calls were too expensive for most of the remote communities invoked to justify the project. Meta later solved much of this accessibility problem. Advertising revenue concentrated in rich consumer markets helps finance global digital infrastructure, while WhatsApp provides a free communications service that functions as a quasi-public utility across countries including India and Brazil. This arrangement can be understood as cross-border financing of locally created consumer surplus. It should not automatically be described as a literal transfer from rich consumers to poor users. The incidence of advertising expenditure is distributed among consumers, firms, workers, suppliers, and shareholders. Low-revenue users also contribute network effects, strategic position, data, and future commercial opportunity. Establishing a true operating subsidy requires a counterfactual: would the platform reduce or withdraw service in the lower-revenue market if revenue from richer markets disappeared? Even where a foreign platform creates enormous local consumer surplus, users may have little control over its ownership, rules, continuity, or technical development. Consumer surplus is therefore distinct from sovereignty. Cloud-based AI could reproduce this structure. Enterprises and consumers in high-wage markets may pay heavily for systems that augment or replace expensive labor. Those revenues could support cheap or free AI services elsewhere, creating broad access while leaving control with a small number of foreign providers. AI may also permit a different path. Frontier models can potentially serve as teachers for smaller, specialized models that operate on local devices or regional infrastructure. If such models are reliable, affordable, legally transferable, modifiable, and maintainable, AI capability could become a locally possessed productive asset rather than a continuously rented service. Rich-market demand would then finance not only global consumption of intelligence, but some diffusion of ownership over intelligent productive capacity. This outcome is plausible but unproven. Written knowledge is incomplete; small models may fail on unusual or high-stakes cases; local inference still requires hardware, electricity, data, maintenance, evaluation, security, and institutional support. Proprietary licenses, locked hardware, controlled update channels, scarce validation capacity, or cloud escalation may preserve substantial dependency. Jeffrey Ding's work on general-purpose technologies adds a further qualification. Possessing an artifact is not equivalent to diffusing a technology throughout an economy. Durable gains require engineers, organizations, complementary infrastructure, trusted institutions, operational data, capital, and the ability to redesign workflows. The central question is therefore not only who can access AI, but who can own, adapt, validate, deploy, and compound it. --- # 2. Central thesis and subsidiary claims ## 2.1 Central thesis **Thesis:** Rich markets may finance AI capabilities that create large consumer and productivity gains elsewhere. Whether this process distributes power rather than merely access depends on the technical form of deployment, ownership and governance arrangements, and the recipient society's capacity to adapt and diffuse the technology. ### Classification - **Type:** synthesis and forward-looking hypothesis - **Confidence:** medium - **Reason for uncertainty:** the economics, capabilities, and governance of cloud and edge AI are changing rapidly; the relevant cross-country evidence is not yet mature. ## 2.2 Subsidiary claims ### Claim A — Universalist rhetoric often outruns practical accessibility A technology's theoretical ability to reach a population does not establish that the population can afford, use, trust, or benefit from it. - **Classification:** historically grounded interpretation - **Confidence:** high - **Illustrative case:** Iridium could technically reach remote users but was initially priced and designed for a narrow affluent market. ### Claim B — Global platforms can separate the geography of financing from the geography of benefit Revenue concentrated in rich markets can help sustain services that create large benefits in low-monetization markets. - **Classification:** strong inference from platform economics - **Confidence:** high in general; medium for claims about the magnitude of any particular country subsidy - **Important qualification:** shared fixed costs, low marginal costs, network effects, and strategic value complicate geographic profit allocation. ### Claim C — Advertising ultimately depends on consumer demand, but its incidence is not borne only by consumers Advertisers pay platforms because they expect advertising to influence economically valuable behavior. Consumer spending replenishes the commercial revenues from which advertising budgets are funded. However, the economic burden of advertising expenditure may be shared among consumers, shareholders, workers, suppliers, and other parties. - **Classification:** standard economic mechanism with incidence qualification - **Confidence:** high ### Claim D — Large consumer surplus does not establish local sovereignty Users can derive substantial value from a service while lacking ownership, portability, governance rights, continuity guarantees, or the ability to modify and independently operate it. - **Classification:** conceptual distinction - **Confidence:** high ### Claim E — Local models could convert AI from a service into a capital good If model weights and necessary complements can be possessed and operated locally, some AI capabilities may become durable productive assets rather than continuously metered cloud services. - **Classification:** technically plausible hypothesis - **Confidence:** medium - **Important qualification:** local execution alone does not establish legal, technical, or institutional sovereignty. ### Claim F — Ownership does not guarantee economy-wide diffusion The impact of a general-purpose technology depends on complementary skills, institutions, infrastructure, organizational redesign, and adaptation throughout the economy. - **Classification:** historically grounded claim - **Confidence:** high - **Key intellectual reference:** Jeffrey Ding, *Technology and the Rise of Great Powers*. ### Claim G — AI may partially accelerate its own diffusion AI can generate code, translate interfaces, create training material, and assist less-experienced implementers, potentially reducing some technical barriers to adoption. - **Classification:** emerging inference - **Confidence:** medium - **Important qualification:** AI cannot by itself resolve political authority, liability, procurement, organizational resistance, physical infrastructure, or institutional trust. --- # 3. Terminology ## 3.1 Access The practical ability to use an AI system or digital service. Access can depend on price, connectivity, hardware, language, literacy, identity requirements, regulation, disability accommodations, and availability in a jurisdiction. Access is broader than theoretical reach and narrower than ownership. ## 3.2 Technical reach The engineering capacity of a system to serve a location or population under ideal conditions. Satellite coverage is an example. Technical reach does not imply affordability, usability, distribution, trust, or institutional integration. ## 3.3 Consumer surplus The difference between the maximum amount a consumer would be willing to pay for a good or service and the amount actually paid. For a user \(i\): $$ CS_i = WTP_i - P_i $$ where \(WTP_i\) is willingness to pay and \(P_i\) is the price paid. For a zero-price digital service, \(P_i = 0\), but this does not mean the total cost to the user is zero. Relevant non-price costs may include attention, privacy loss, data provision, exposure to advertising, switching costs, security risks, and dependency. ## 3.4 Cross-border consumer-surplus transfer A useful but nonstandard shorthand for a system in which economic activity or revenue concentrated in one set of countries helps support services that create substantial consumer surplus in another. The term is rhetorically useful but economically imprecise. Consumer surplus itself is not normally transferred like cash. Financing, technology, software, knowledge, or access crosses borders; surplus is then created locally. ### Preferred precise formulation **Cross-border financing of locally consumed digital surplus through a globally integrated platform.** ## 3.5 Cross-border operating subsidy A situation in which revenue attributable to one country or user group does not cover the avoidable cost of serving it, and revenue earned elsewhere enables that service to continue. A strong test is: $$ R_L < C_L^{avoidable} $$ where \(R_L\) is revenue attributable to the lower-monetization market and \(C_L^{avoidable}\) is the cost the provider would avoid by no longer serving it. This is stronger than showing that the market fails to cover an allocated share of global fixed costs. ## 3.6 Shared-fixed-cost spillover A lower-income market benefits from a technical asset largely financed by demand elsewhere, while the marginal cost of extending that asset to the additional market is small. The recipient market may still be contribution-positive even if it could not independently finance development of the system. This pattern is common in software and knowledge goods. ## 3.7 Cross-side subsidy In a multisided market, a platform charges one participant group less because its participation makes another group more valuable. Advertiser-funded media and zero-price digital platforms are standard examples. A cross-side subsidy need not be cross-border. A global platform can combine both structures. ## 3.8 Network barter A descriptive term for an implicit exchange in which one population contributes revenue while another contributes network scale, social connections, content, data, strategic reach, or future option value. This framing cautions against treating low-revenue users as passive recipients of charity. ## 3.9 Platform welfare Consumer benefit delivered through discretionary access to a privately owned platform. The service may be highly valuable or free at the point of use, but recipients do not necessarily acquire ownership, governance rights, interoperability, or a durable entitlement. This is not an established term of art; it is used here as an analytical contrast with technology transfer. ## 3.10 Technology transfer The movement of productive knowledge, tools, intellectual property, operational capability, and know-how from one organization or jurisdiction to another. Strong technology transfer leaves recipients more capable of operating, modifying, maintaining, and reproducing the relevant system. ## 3.11 Artifact diffusion Distribution of the technological artifact itself: model weights, software, hardware designs, documentation, or related tools. Artifact diffusion is not equivalent to effective adoption. ## 3.12 Systemic diffusion Widespread adaptation of a technology across firms, public institutions, sectors, and workflows, accompanied by the complementary skills and organizational changes required to generate sustained productivity gains. ## 3.13 Diffusion capacity A society's ability to evaluate, absorb, adapt, deploy, maintain, and improve a technology. Relevant complements include: - technical and managerial skills; - universities and vocational training; - operational and scientific data; - reliable infrastructure; - access to capital; - procurement competence; - standards and certification; - organizational flexibility; - competition and entrepreneurial capacity; - trusted mechanisms for liability and redress. ## 3.14 AI sovereignty The practical capacity of a person, institution, or polity to continue using, governing, modifying, replacing, and independently operating an AI system without unilateral dependence on an external provider. Sovereignty is not binary. It can exist at several layers: - model weights; - compute hardware; - software stack; - data and knowledge bases; - identity and payment systems; - update and security channels; - evaluation and certification; - cloud escalation; - legal authority. ## 3.15 Small language model (SLM) A language model designed to use fewer parameters and less compute than frontier-scale models. The category has no stable numerical boundary. In this companion, the important property is not model size alone but whether the model can be economically deployed on local devices or regional infrastructure for a bounded task. ## 3.16 Distillation A training process in which a smaller model learns from the outputs, representations, or demonstrated behavior of a larger teacher model. Distillation can transfer selected capabilities, but it does not guarantee equivalent performance, robustness, calibration, or coverage of unusual cases. ## 3.17 Edge deployment Execution on user-controlled or locally controlled hardware, including phones, personal computers, institutional servers, vehicles, sensors, and regional appliances, rather than on a distant provider's cloud. Edge deployment can improve latency, privacy, continuity, and marginal cost. It does not by itself establish ownership or sovereignty. --- # 4. Analytical framework ## 4.1 Separate six questions Analysis of global AI diffusion should distinguish: 1. **Who finances initial development?** 2. **Who pays for ongoing operation?** 3. **Who receives consumer or productivity benefits?** 4. **Who captures revenue, profit, wages, tax receipts, and technical learning?** 5. **Who governs access and operating rules?** 6. **Who accumulates the capacity to adapt and reproduce the system?** These groups need not be located in the same country. ## 4.2 A broader national-welfare frame Consumer surplus is only one component of national welfare. A fuller conceptual accounting is: $$ W_j = CS_j + PS_j + L_j + T_j + K_j - E_j - D_j $$ where: - \(CS_j\): consumer surplus in country \(j\); - \(PS_j\): locally captured producer surplus; - \(L_j\): labor income and employment effects; - \(T_j\): tax and fiscal effects; - \(K_j\): accumulated knowledge, institutional capacity, and technical capability; - \(E_j\): social, political, environmental, and security externalities; - \(D_j\): dependency, switching, and loss-of-control costs. This is an analytical schema, not an immediately estimable equation. Its purpose is to prevent high consumer surplus from being mistaken for comprehensive development or national power. ## 4.3 Four stages of AI diffusion ### Stage 1 — Access People can use an AI system at an affordable financial and practical cost. **Primary outcome:** consumption benefit. ### Stage 2 — Possession Local actors can run the model or system on infrastructure they control and can continue operating it without permission for each use. **Primary outcome:** operational autonomy. ### Stage 3 — Adaptation Local actors can modify the system for local languages, laws, workflows, knowledge, risks, and material conditions. **Primary outcome:** technical and institutional capability. ### Stage 4 — Systemic diffusion Firms, farms, clinics, schools, and governments reorganize around the technology and translate it into widespread productivity and welfare gains. **Primary outcome:** economy-wide transformation. Each stage is generally more demanding than the previous one. Progress is not automatic or irreversible. ## 4.4 Sovereignty test for locally deployed AI A community or institution has stronger practical sovereignty when it can: 1. possess and run the necessary weights without recurring external authorization; 2. operate on hardware it controls or can replace; 3. inspect model provenance and limitations; 4. modify or replace local knowledge and policy layers; 5. retain sensitive usage data locally; 6. continue operating if the original provider disappears; 7. choose whether and when to update; 8. procure interoperable alternatives; 9. commission independent evaluation; 10. train local personnel to maintain the system; 11. escalate difficult cases without being locked to one supplier; 12. legally redistribute or reproduce the system where appropriate. Few real systems will satisfy all twelve conditions. The test is intended to identify the location and degree of dependency. ## 4.5 Evidence labels used in this document - **Documented fact:** directly supported by a cited primary source or strong empirical study. - **Literature-backed interpretation:** consistent with a recognized body of scholarship but not uniquely established by it. - **Inference:** a reasoned conclusion from documented facts. - **Hypothesis:** a testable claim for which evidence is incomplete. - **Speculation:** a plausible future path with substantial uncertainty. - **Contested claim:** a claim for which credible evidence or theory points in different directions. --- # 5. Questions this companion will examine 1. Is "cross-border consumer-surplus transfer" a useful concept, and what more precise mechanisms sit beneath it? 2. Does Meta's global portfolio contain a true geographic operating subsidy, a shared-fixed-cost spillover, a network barter, or some combination? 3. What can international telephone settlements, universal-service systems, advertising-supported media, innovation economics, and global-public-good theory teach us? 4. How does the creation of consumer surplus differ from local value capture and sovereignty? 5. Under what technical and legal conditions can AI become a locally possessed capital good? 6. Which tasks are plausible candidates for edge or regional SLM deployment? 7. Which complements remain scarce even if model weights and inference become cheap? 8. How does Jeffrey Ding's diffusion framework alter the analysis? 9. Could AI lower its own barriers to diffusion? 10. Which subsidy designs purchase temporary access, and which create transferable capacity? 11. What evidence would confirm or falsify the main thesis? 12. What are the strongest objections to the argument? --- # Part II — Prior art and historical foundations --- # 6. Prior-art map "Cross-border consumer-surplus transfer" is not an established term of art. The phenomenon sits at the intersection of several literatures that use different units of analysis and answer different questions. | Literature | Principal question | Contribution to this inquiry | Principal limitation | |---|---|---|---| | Two-sided markets | How should a platform price interdependent participant groups? | Explains why one side may receive free or subsidized service because it creates value for another | Usually abstracts from geography, national welfare, and sovereignty | | Digital consumer welfare | How much value do people derive from zero-price digital goods? | Provides methods and cross-country estimates of consumer surplus | Does not establish which market financed the benefit or who bears the cost | | Innovation appropriability | How much of an innovation's social return can its producer capture? | Explains why consumers and imitators may receive most of the value created by innovation | Often domestic and aggregate; says little about international distribution or control | | International knowledge diffusion | How do ideas developed in one country affect productivity and welfare elsewhere? | Models cross-border gains from foreign innovation and adaptation | Usually focuses on production, trade, patents, and productivity rather than zero-price services | | Global public goods | When is knowledge non-rival and difficult to exclude? | Clarifies why knowledge can spread globally at low marginal cost and why it may be underproduced | The cost of transmission, adaptation, validation, and use can remain high | | International telecom settlements | How should carriers compensate one another for cross-border traffic? | Provides a literal example of large rich-to-poor financial flows connected to communications | Transfers to carriers did not necessarily become consumer welfare or infrastructure investment | | Universal service | How should profitable users or regions support uneconomic coverage? | Supplies cost tests, entitlement concepts, and models of geographic cross-subsidy | Usually operates within one polity and under public mandate | | Advertising-supported media | Why can audiences consume content below cost or for free? | Provides a historical model of indirect financing by advertisers | Advertisers generally seek the same audience receiving the service; interpersonal infrastructure is different from media content | | Digital development | Who creates and captures value from digitalization? | Separates access and use from profits, taxes, jobs, learning, and productive capacity | Often lacks direct measures of consumer surplus | | Digital colonialism and sovereignty | How can useful foreign infrastructure generate dependency and asymmetric power? | Adds ownership, standards, data, governance, and exit to welfare analysis | Sometimes treats extraction as self-evident and underweights genuine user benefit | | Technology diffusion and state capacity | Why do some societies translate general-purpose technologies into power and productivity? | Shows that possession of an artifact is not equivalent to systemic adoption | Historical analogies may not capture AI's speed or its ability to assist its own deployment | No single literature captures the entire object. A useful synthesis must keep four separate ledgers: 1. **Financing:** where revenue, capital, and public or philanthropic subsidy originate. 2. **Welfare:** where consumer and producer benefits appear. 3. **Value capture:** where profits, wages, taxes, data, and organizational learning accumulate. 4. **Control:** who can govern, modify, replace, or terminate the infrastructure. --- # 7. Two-sided markets and cross-side subsidy ## 7.1 Core mechanism A two-sided or multisided platform enables interactions among participant groups whose demand is interdependent. Credit-card networks connect cardholders and merchants; media businesses connect audiences and advertisers; marketplaces connect buyers and sellers. The foundational work of Jean-Charles Rochet and Jean Tirole emphasizes that the *structure* of prices across sides matters, not only the total price collected. Mark Armstrong similarly shows that equilibrium prices depend on cross-group externalities, charging structures, and whether users participate on one platform or several. A stylized platform chooses prices \(p_U\) and \(p_A\) for users and advertisers: $$ \Pi = p_U N_U + p_A N_A - C(N_U,N_A) $$ where each side's participation depends partly on participation by the other. If an additional user substantially increases advertiser value, the platform may rationally set: $$ p_U \leq MC_U $$ and recover revenue from advertisers. In some configurations the user price may be zero or negative. ## 7.2 Application to Meta This explains why zero-price access is not anomalous and why price need not track the cost of serving each user. Meta does not need every user to generate enough directly attributable revenue to cover an allocated share of global costs. A user may contribute value by: - supplying attention and advertising inventory; - completing a social graph; - producing or forwarding content; - attracting other users; - making a communication standard more universal; - strengthening the platform's position against competitors; - supplying future commercial option value. The global extension is not merely "advertisers subsidize users." Different national populations can occupy different economic roles. High-income markets may supply most revenue, while lower-income markets supply scale, network completeness, strategic reach, and significant user welfare. ## 7.3 Why the word subsidy can mislead A cross-side subsidy is a pricing description, not necessarily a gift. The nominally subsidized side may provide indispensable non-cash inputs to the system. In a nightclub analogy, one group may be admitted cheaply because its presence attracts a higher-paying group; it is being priced strategically, not supported philanthropically. The same caution applies geographically. Indian WhatsApp users are not simply costs borne by American users. They help make WhatsApp a global standard, connect international families and businesses, deter competitors, and preserve potential future monetization. ## 7.4 What the literature leaves unresolved Two-sided-market theory alone does not tell us: - whether India is contribution-negative for Meta; - how shared R&D and infrastructure costs should be geographically allocated; - who ultimately bears advertising expenditure; - how much consumer surplus users receive; - whether global access creates or reduces national productive capacity; - whether foreign corporate control creates material dependency costs. **Evidence status:** The general pricing mechanism is well established. Its precise geographic incidence inside Meta is not publicly observable. ### Key sources - Jean-Charles Rochet and Jean Tirole, “Platform Competition in Two-Sided Markets,” *Journal of the European Economic Association* 1, no. 4 (2003): 990–1029. https://doi.org/10.1162/154247603322493212 - Mark Armstrong, “Competition in Two-Sided Markets,” *RAND Journal of Economics* 37, no. 3 (2006): 668–691. https://doi.org/10.1111/j.1756-2171.2006.tb00037.x --- # 8. Measuring welfare from zero-price digital goods ## 8.1 The measurement problem National accounts value marketed production, not the full benefit users receive. A zero-price digital service can create large welfare gains while making little direct contribution to measured household consumption. One approach uses incentivized choice experiments to estimate a user's willingness to accept compensation in exchange for temporarily losing access. If a user would require \(WTA_i\) to surrender a free service for a defined period, researchers treat this as evidence about the service's consumer value. This work does not imply that stated consumer value should simply be added to GDP. It addresses a different question: how much welfare may be omitted when digital services have a zero monetary price? ## 8.2 *The Digital Welfare of Nations* Brynjolfsson, Collis, Liaqat and coauthors conducted incentivized online choice experiments with representative samples totaling nearly 40,000 people across 13 countries. For ten popular digital goods, they estimated more than $2.5 trillion in annual aggregate consumer welfare, roughly 6% of the participating countries' combined GDP. They also found that lower-income individuals and countries received larger gains relative to income than richer ones. This is unusually relevant to the present thesis because it documents a geographical pattern consistent with welfare equalization: digital goods can matter proportionately more in poorer countries even when provider revenue and ownership are concentrated in richer ones. ## 8.3 What the study does not prove The results do **not** establish that rich-country users financed poor-country users. They measure benefits, not the causal incidence of costs or revenues. They also require several qualifications: 1. **Willingness to accept may capture dependency.** A high valuation can reflect the absence of interoperable alternatives or the fact that a person's entire network uses one platform. 2. **Monetary measures vary with income.** Absolute valuations by poorer users are constrained by economic circumstances even when a service is central to their lives. 3. **Private benefit is not social benefit.** Individual valuations may not net out misinformation, fraud, polarization, privacy loss, or effects on non-users. 4. **Zero price is not zero cost.** Users may supply attention, data, content, and exposure to commercial persuasion. 5. **Consumer welfare is not productive capacity.** A valuable messaging tool may improve daily life without creating domestic firms, tax revenue, technical learning, or strategic autonomy. ## 8.4 Disclosure and interpretation The NBER page discloses that several authors were Meta employees with financial interests in the company, one was a former employee, and other authors had previously received unrestricted gifts from Meta or worked at an institution partly supported by varied funders. These disclosures do not invalidate the study. They do increase the importance of methodological scrutiny and independent replication because the finding aligns closely with a major platform's public account of its social value. **Evidence status:** Strong evidence that zero-price digital goods create large measured consumer welfare and may yield proportionately larger benefits in lower-income countries. No direct evidence of a cross-border operating subsidy. ### Key source - Erik Brynjolfsson et al., “The Digital Welfare of Nations: New Measures of Welfare Gains and Inequality,” NBER Working Paper 31670 (2023). https://doi.org/10.3386/w31670 --- # 9. Innovation appropriability: creators capture only part of the surplus ## 9.1 Social and private returns Innovation can lower production costs, improve quality, create new products, and enable follow-on invention. The producing firm may capture some value through price, intellectual property, secrecy, brand, distribution, and first-mover advantage. Competition, imitation, expiration of rights, and complementary innovation pass much of the remaining value to consumers and other producers. A simplified decomposition is: $$ SV = \Pi_I + CS + PS_O + S $$ where: - \(SV\) is total social value; - \(\Pi_I\) is profit captured by the innovator; - \(CS\) is consumer surplus; - \(PS_O\) is producer surplus captured by other firms; - \(S\) represents wider spillovers. ## 9.2 Nordhaus's estimate William Nordhaus examined Schumpeterian profits in the US non-farm business economy over 1948–2001. His conclusion was that producers captured only a small fraction of the social returns to technological advances, with most benefits passed to consumers rather than retained as producer profits. A widely cited interpretation of his estimates places the innovator share near 2.2%. That number should not be treated as a universal constant. It reflects a particular model, period, economy, definition of excess innovative profit, and set of measurement assumptions. Platform monopolies, pharmaceutical patents, frontier AI, and open-source software may exhibit very different appropriability. The durable point is conceptual: > Financing innovation and ultimately receiving its surplus are separate events. ## 9.3 International extension Once an innovation crosses borders, consumers and producers in countries that did not finance the original R&D may capture substantial benefits. This can occur through: - imported capital goods; - lower-priced products; - trade and foreign direct investment; - patents and scientific publication; - worker mobility; - imitation and reverse engineering; - open-source software; - standards and protocols; - model weights, distillation, and synthetic training data. The global distribution of surplus then depends on excludability, licensing, complementary assets, competition, and absorptive capacity. ## 9.4 Relevance to AI Rich-country firms may finance frontier training because AI can substitute for expensive labor or create valuable products in those markets. If model capability is subsequently available worldwide at low marginal cost, users and firms elsewhere may capture value far beyond their contribution to initial development. This does not by itself imply redistribution in a political or fiscal sense. It may instead be the ordinary pattern of innovation surplus amplified by software's low reproduction costs. **Evidence status:** Strong historical and theoretical support for incomplete appropriability. The size and geographic distribution of AI's eventual surplus remain unknown. ### Key sources - William D. Nordhaus, “Schumpeterian Profits in the American Economy: Theory and Measurement,” NBER Working Paper 10433 (2004). https://doi.org/10.3386/w10433 - Kenneth J. Arrow, “Economic Welfare and the Allocation of Resources for Invention,” in *The Rate and Direction of Inventive Activity* (1962). - Richard R. Nelson, “The Simple Economics of Basic Scientific Research,” *Journal of Political Economy* 67, no. 3 (1959): 297–306. --- # 10. International knowledge diffusion ## 10.1 Ideas cross borders through multiple channels The international growth and trade literature treats foreign knowledge as an input into domestic productivity and innovation. Diffusion can occur through trade, patent disclosure, scientific networks, foreign direct investment, worker movement, multinational supply chains, and observation of foreign products and processes. Cai, Li, and Santacreu model innovation and knowledge diffusion across countries and sectors. Their quantitative framework finds that foreign knowledge spillovers can amplify dynamic welfare gains from trade. In their counterfactual analysis, smaller countries with less R&D have more to gain from access to foreign pools of ideas. ## 10.2 Access to an idea is not equivalent to adoption Knowledge diffusion models often require or imply some capacity to recognize, absorb, and use foreign ideas. Related work on absorptive capacity emphasizes that prior knowledge, human capital, institutions, and complementary R&D affect whether a recipient can exploit external knowledge. This produces a conditional rather than automatic convergence mechanism: $$ \text{Foreign knowledge access} \times \text{absorptive capacity} \rightarrow \text{domestic productivity gain} $$ If absorptive capacity is weak, formally available knowledge may create little economic change. ## 10.3 Relevance to AI and SLMs Open or transferable model weights could dramatically reduce the cost of acquiring a technological artifact. But realized gains may still depend on: - local-language and domain data; - hardware and systems engineering; - trustworthy evaluation; - institutional authority to change workflows; - credit and procurement; - integration with physical supply chains; - maintenance and security; - domain experts capable of recognizing failure. AI may reduce some of these requirements by assisting with programming, translation, training, and diagnosis. Whether it reduces them enough to change historical diffusion patterns is an open empirical question. **Evidence status:** Strong support for economically important international knowledge diffusion and heterogeneous gains. Application to downloadable AI capability is an informed analogy, not an established result. ### Key source - Jie Cai, Nan Li, and Ana Maria Santacreu, “Knowledge Diffusion, Trade and Innovation across Countries and Sectors” (working paper, 2017). https://economics.yale.edu/sites/default/files/cls-2017-santacreu.pdf --- # 11. Knowledge as a global public good ## 11.1 Non-rivalry and excludability Joseph Stiglitz describes knowledge as a global public good because one person's use of an idea does not ordinarily prevent another's, and the idea may remain useful across borders. A mathematical theorem can be used in many countries without being depleted. The central economic tension is: - efficient use favors a price near the marginal cost of an additional user, often near zero; - production requires resources and therefore some financing mechanism. Intellectual-property rights, secrecy, public research, prizes, philanthropy, and complementary-service business models are alternative ways of resolving this tension. ## 11.2 Transmission and use still cost money The public-good quality belongs to the knowledge, not necessarily to every system required to deliver or apply it. Even if a model's encoded knowledge can be copied cheaply, practical use may require: - storage and compute; - electricity; - networking and updates; - sensors and current data; - user interfaces; - local adaptation; - certification and liability systems; - skilled maintenance. Therefore: $$ MC_{copy} \approx 0 $$ can coexist with: $$ C_{effective\ deployment} > 0 $$ ## 11.3 Artificial excludability Model capability may be technically reproducible but institutionally enclosed through: - API-only provision; - secret weights and training methods; - copyright, contract, and trade-secret law; - licenses restricting redistribution or particular uses; - controlled app stores and hardware; - remote authentication; - unavailable evaluation or fine-tuning tools; - proprietary data and retrieval systems. Reichman and Maskus describe a related tension as the globalization of private knowledge goods and the privatization of global public goods. Strong global intellectual-property rules may support innovation and formal technology transfer, but they may also obstruct competition, follow-on innovation, public health, education, and scientific advance in developing countries. ## 11.4 Implication for the essay AI's technical character does not determine its institutional form. The same underlying capability could be organized as: - a metered private service; - a licensed local appliance; - an open model with proprietary complements; - a publicly maintained digital utility; - a reproducible cognitive commons. The relevant policy question is not whether AI *is* a public good in the abstract. It is which layers are non-rival, which remain scarce, and where excludability is technically necessary, commercially imposed, or legally constructed. **Evidence status:** The public-good analysis of knowledge is established. Treating trained AI capability as knowledge requires care because inference consumes real compute and because weights do not contain every complement required for reliable use. ### Key sources - Joseph E. Stiglitz, “Knowledge as a Global Public Good,” in *Global Public Goods* (Oxford University Press, 1999). - Jerome H. Reichman and Keith E. Maskus, “The Globalization of Private Knowledge Goods and the Privatization of Global Public Goods,” *Journal of International Economic Law* 7 (2004/2005): 279–320. https://scholarship.law.duke.edu/faculty_scholarship/2144/ --- # 12. International telephone settlements ## 12.1 The accounting-rate system Under the traditional international telephone regime, the carrier originating a call compensated the carrier terminating it in another country. Where traffic was imbalanced, the originating carrier made a net settlement payment based on excess minutes and a bilaterally negotiated accounting rate. Because more traffic often originated in developed countries than in developing countries, and because settlement rates could exceed the incremental cost of termination, substantial funds flowed from developed-country carriers to developing-country carriers. The International Telecommunication Union estimated that net settlement flows from developed to developing countries totaled about $40 billion between 1993 and 1998. ## 12.2 Why this is the closest literal precedent Unlike the platform case, this was an observable cross-border payment connected directly to communications traffic: ```text Caller payment → originating carrier → settlement payment → foreign terminating carrier ``` Developing countries and incumbent carriers argued that settlement revenue financed telecommunications investment and broader domestic service. ## 12.3 Transfer to an incumbent is not necessarily consumer benefit Scott Wallsten examined data for 179 countries from 1985 through 1998. He reported that developing countries received about $35 billion in net payments from US carriers during that period, but found no evidence that the payments increased telephone penetration or telecommunications-equipment imports. Higher settlement rates were negatively correlated with international traffic. The case therefore separates three things often conflated in subsidy rhetoric: 1. a cross-border cash transfer; 2. revenue received by a domestic or state-owned provider; 3. welfare or infrastructure gains for the intended population. A payment can be absorbed as monopoly rent, fiscal revenue, patronage, inefficiency, or profit rather than translated into universal service. ## 12.4 Iridium's political economy Iridium threatened to route communications around terrestrial networks and national international gateways. Yet deployment required licenses, spectrum coordination, country agreements, and political support. Motorola therefore had incentives to restore national intermediaries contractually through local partners, gateways, revenue-sharing arrangements, and other concessions. The broader lesson is that a technology can eliminate an intermediary technically while reproducing it economically or politically. A borderless network may still need to distribute rents created by borders. This companion treats detailed claims about particular Iridium agreements as requiring verification against John Bloom's source notes, contemporaneous agreements, or other primary documentation before publication as fact. ## 12.5 From settlements to internet applications Internet-based calling weakened the old settlement architecture. A WhatsApp call ordinarily travels as data purchased through local internet access rather than as a specially priced international telephone call. The local carrier may earn data revenue, but it does not necessarily receive a bilateral international-call settlement. Meta internalizes a different financing structure: ```text High-value commercial activity and advertising markets → Meta's consolidated revenue → global shared infrastructure → locally consumed WhatsApp service ``` The transfer becomes less visible, less treaty-based, and less accountable to recipient states. **Evidence status:** The scale and mechanics of the historical settlement system are well documented. Its use as an analogy for Meta and AI is interpretive. ### Key sources - International Telecommunication Union, “Accounting Rate Reform undertaken by ITU-T Study Group 3.” https://www.itu.int/en/ITU-T/studygroups/2013-2016/03/Pages/accounting-rate.aspx - Scott J. Wallsten, “Telecommunications Investment and Traffic in Developing Countries: The Effects of International Settlement Rate Reforms,” *Journal of Regulatory Economics* 20 (2001): 307–323. https://doi.org/10.1023/A:1011171110899 - John Bloom, *Eccentric Orbits: The Iridium Story* (Grove Press, 2016). --- # 13. Universal service and geographic cross-subsidy ## 13.1 The public-service rationale Postal, telephone, electricity, transport, and other network industries often serve high-cost locations at prices below their fully allocated cost. Dense or commercially attractive users may support remote, rural, or low-income users because universal access creates social and network benefits. This resembles the WhatsApp and prospective AI cases, but traditional universal service normally has three properties absent from discretionary platform provision: 1. an explicit public objective; 2. a legal or regulatory mandate; 3. an identified financing mechanism. ## 13.2 Cost tests A claim of cross-subsidy depends on the cost concept used. ### Incremental or avoidable-cost test A population is strongly subsidized if attributable revenue is below the costs that would disappear if service ended: $$ R_g < C_g^{avoidable} $$ ### Stand-alone-cost test A population may benefit from shared infrastructure if it could not economically reproduce the service alone: $$ R_g < C_g^{standalone} $$ This does not establish that serving it reduces the provider's profit. Most network users benefit from shared fixed costs. ### Fully allocated-cost test Analysts may assign global overhead and R&D across regions. Results depend heavily on the allocation rule and can make almost any low-revenue market appear subsidized. For digital platforms, avoidable cost is usually the cleanest economic test, while stand-alone cost is the most revealing test of dependency. ## 13.3 Entitlement versus discretion A user of a regulated universal service may possess a legal claim to access, continuity, nondiscrimination, or affordability. A user of a free global platform usually possesses contractual permission that the provider can modify, subject to applicable law and political constraints. This yields an important distinction: > A service can function socially as a public utility without being governed institutionally as one. ## 13.4 Quality stratification Universal access can coexist with unequal quality. AI may be formally available everywhere while differing in: - model capability; - latency and usage limits; - language performance; - privacy and data retention; - uptime guarantees; - tool and database access; - human escalation; - liability and redress. The future inequality may therefore sit inside the service rather than between those who have and lack it. **Evidence status:** Geographic cross-subsidy and universal-service principles are established in network economics. Application to global privately governed AI requires new institutional analysis. --- # 14. Advertising-supported media and the payer chain ## 14.1 Historical structure Advertising has long financed newspapers, magazines, commercial radio, television, search, social media, and other services at prices below the cost of production or at a zero monetary price. A simplified chain is: ```text Audience attention and expected purchasing behavior → advertiser demand → media or platform revenue → subsidized content or service ``` The audience is not merely a beneficiary. Its attention and purchasing potential are inputs sold to advertisers. ## 14.2 From advertiser to consumer The statement "advertisers pay" stops the incidence analysis too early. Firms advertise because they expect an economic return through some combination of higher sales, stronger pricing power, customer retention, or protection of market position. Consumer demand makes that expenditure rational. The intuitive Meta chain is: ```text Affluent consumer demand and spending → commercial value to advertisers → advertising budgets → Meta revenue → global infrastructure and services ``` However, it is too strong to say that consumers alone pay for advertising. Depending on competition and market structure, advertising expenditure may reduce shareholder profits, affect wages or supplier payments, or be reflected in consumer prices. Advertising can also improve matching and competition rather than acting only as a cost embedded in products. The essay therefore uses the careful formulation that Western consumer spending **indirectly finances** the service. ## 14.3 What is distinctive about Meta's portfolio Traditional advertising generally finances content delivered to the audience being monetized. Meta adds two separations: 1. **Geographic separation:** the most valuable advertising audiences and the largest user populations may be in different countries. 2. **Product separation:** advertising on Facebook and Instagram can finance WhatsApp, where the recipient user may see no advertisement. This is closer to a conglomerate-level cross-subsidy than ordinary ad-supported media. ## 14.4 Implication for AI Free AI could be financed through several commercial channels: - advertising displayed alongside answers; - sponsored recommendations; - commissions on agent-executed transactions; - lead generation for lenders, insurers, healthcare providers, or merchants; - enterprise subscriptions subsidizing consumer access; - premium users subsidizing free tiers; - one product in a corporate portfolio financing another. AI introduces a sharper conflict than conventional advertising because an assistant may not merely display persuasion. It may interpret intent, recommend an action, and execute a purchase. A free agricultural adviser funded by seed, fertilizer, insurance, or credit providers could be useful while also being commercially compromised. The critical governance question becomes: > Is the assistant acting for the user, for the party paying for access to the user, or under a disclosed and enforceable balance between the two? **Evidence status:** Advertising support of zero-price media is well established. The likely commercial architecture of AI assistants remains speculative and will vary by provider and jurisdiction. --- # 15. Digital development: value creation is not value capture ## 15.1 National benefit has several components A foreign digital service can create substantial consumer welfare while its producer captures most revenue, profit, technical learning, and strategic control abroad. Developing countries may contribute users, data, labor, cultural material, and market growth while collecting little tax and building few domestic firms. UNCTAD's *Digital Economy Report 2019* frames this problem as one of value creation and value capture. The report emphasizes the highly concentrated geography of major digital platforms and the development implications of data-driven business models. A complete national accounting should ask where the following accrue: - consumer surplus; - producer surplus; - wages and high-skill employment; - tax revenue; - data and operational learning; - intellectual property; - standards-setting influence; - entrepreneurial spillovers; - bargaining power; - continuity and security risks. ## 15.2 High consumer welfare can coexist with weak development capture There is no contradiction in all of the following being true: - WhatsApp is enormously useful to Indian households and firms; - Meta earns little direct revenue per Indian user compared with richer markets; - Indian participation contributes strategic value to Meta; - most platform profits and core technical capabilities accrue outside India; - local communication becomes dependent on a foreign-controlled platform. The same multidimensional outcome is possible for AI. ## 15.3 From consumption to production Development gains are more durable when foreign technology supports: - domestic complements and suppliers; - local technical and managerial learning; - new firms and export capabilities; - public-sector competence; - local data stewardship; - adaptation to domestic needs; - credible substitution among providers. Locally operable SLMs could help, but only if they form part of a broader process of capability accumulation. **Evidence status:** Strong conceptual and descriptive support. Measuring dependency and capability accumulation remains difficult. ### Key source - United Nations Conference on Trade and Development, *Digital Economy Report 2019: Value Creation and Capture—Implications for Developing Countries*. https://unctad.org/system/files/official-document/der2019_en.pdf --- # 16. Digital colonialism and sovereignty ## 16.1 The critique The digital-colonialism literature argues that primarily US-based technology companies can occupy roles analogous in some respects to earlier imperial powers. They design infrastructures around their economic interests, expand them globally, establish strong dependencies, concentrate data and technical control, and rely in part on labor and resources from the Global South. The analogy is strongest when it concerns: - externally controlled infrastructure; - monopolistic or oligopolistic dependence; - asymmetric standard-setting power; - extraction of data or labor under unequal terms; - limited local ability to exit or govern; - displacement of domestic alternatives. ## 16.2 Limits of the colonial analogy The analogy can become imprecise if it treats users as receiving no benefit or collapses every unequal exchange into coercive extraction. Unlike many historical colonial relationships, adoption of a digital platform can be voluntary at the individual level and can create very large consumer surplus. But voluntary use does not eliminate structural dependence. Network effects, institutional integration, lack of substitutes, and collective-action problems can make individual exit possible in theory but costly in practice. ## 16.3 Productive synthesis Consumer economics and sovereignty analysis illuminate different dimensions: | Consumer-welfare lens | Sovereignty lens | |---|---| | How valuable is the service to users? | Who owns and governs the system? | | What price and non-price costs do users bear? | Can users or states modify, replace, or continue it? | | What substitutes are available? | Who sets standards and controls data? | | What harms and externalities should be netted out? | What dependency and bargaining-power effects accumulate? | Neither lens should displace the other. The strongest defensible conclusion is: > A foreign platform can improve welfare substantially while also deepening asymmetric dependency. ## 16.4 Application to AI Cloud AI can create a more consequential dependency than messaging because it may mediate: - administrative decisions; - scientific and technical work; - education; - healthcare guidance; - business operations; - software production; - access to markets and finance; - interpretation of law and policy. Edge deployment can reduce dependence on continuous foreign service but may preserve upstream control through chips, licenses, model updates, evaluation systems, and difficult-case escalation. **Evidence status:** The existence of foreign-platform dependence is well supported descriptively. “Colonialism” is a contested interpretive frame and should not be presented as a neutral empirical classification. ### Key source - “An Intellectual History of Digital Colonialism,” *Journal of Communication* 75, no. 5 (2025): 385–397. https://academic.oup.com/joc/article/75/5/385/8078024 --- # 17. Ding: diffusion, skill infrastructure, and national power ## 17.1 Innovation-centered versus diffusion-centered accounts Jeffrey Ding argues that explanations of technological power often overemphasize breakthrough invention. His alternative centers on the ability of states to adapt and embrace general-purpose technologies at scale. Across historical industrial revolutions, the decisive factors include institutional adaptations and broad skill infrastructure, not merely a small number of frontier laboratories or heroic inventors. ## 17.2 Implication for AI A country may enjoy one or more of the following without achieving systemic transformation: - access to a frontier chatbot; - permission to call an API; - downloadable weights; - a domestic benchmark success; - a small group of elite researchers; - a government-branded national model. Diffusion requires the technology to enter ordinary production and administration. Relevant capabilities include: - large populations of competent engineers and technicians; - managers able to redesign organizations; - domain experts who can supervise and evaluate systems; - educational and vocational institutions; - deployment capital; - reliable infrastructure and operational data; - standards, liability, and procurement systems; - competition and incentives to adopt. ## 17.3 Four distinct surpluses Ding's framework suggests separating: 1. **Consumption surplus:** people receive useful answers or services. 2. **Productivity surplus:** organizations produce more or better output from given inputs. 3. **Capability surplus:** local actors accumulate reusable skills, institutions, data, and technical systems. 4. **Power surplus:** the society gains durable economic, strategic, and bargaining capacity relative to others. Cross-border access can deliver the first. It may contribute to the second. It does not automatically create the third or fourth. ## 17.4 Can AI accelerate its own diffusion? AI may be unusual because it can assist with some complementary tasks: - generating integration code; - translating interfaces and documentation; - creating curricula and training materials; - helping technicians troubleshoot deployments; - converting policy or manuals into machine-usable workflows; - lowering the expertise required for software development. This creates a testable hypothesis: AI could reduce the technical skill threshold for its own adoption and compress diffusion timelines relative to previous general-purpose technologies. The counterargument is that the hardest constraints may be organizational and political rather than technical. AI cannot independently assign legal liability, build trust, authorize procurement, overcome incumbent resistance, provide missing physical goods, or create reliable public institutions. **Evidence status:** Ding's historical diffusion thesis is well developed. AI-assisted self-diffusion is an emerging hypothesis. ### Key source - Jeffrey Ding, *Technology and the Rise of Great Powers: How Diffusion Shapes Economic Competition* (Princeton University Press, 2024). https://press.princeton.edu/books/paperback/9780691260341/technology-and-the-rise-of-great-powers --- # 18. Integrated lessons from the prior art ## 18.1 What is well established 1. Platforms can rationally charge one participant group little or nothing because that group creates value elsewhere. 2. Zero-price digital goods can generate very large consumer welfare that national accounts do not capture well. 3. Innovators often capture only part of the social surplus their inventions create. 4. Knowledge developed in one country can generate productivity and welfare gains elsewhere. 5. Knowledge has non-rival properties, but transmission and effective use remain costly. 6. Cross-border communications systems have previously produced large explicit financial transfers. 7. Transfers to intermediaries do not guarantee benefits to intended populations. 8. Universal access can coexist with unequal service quality. 9. High consumer surplus can coexist with concentrated ownership, profits, technical capacity, and control. 10. General-purpose technologies require complementary skills and institutional adaptation to transform economies. ## 18.2 What remains unproven 1. Whether serving India through WhatsApp is contribution-negative on an avoidable-cost basis. 2. The ultimate incidence of Meta's advertising expenditure across consumers, firms, workers, suppliers, and shareholders. 3. The amount of WhatsApp consumer surplus by country net of non-price costs and externalities. 4. Whether frontier models are close to complete on the written knowledge needed for high-value bounded tasks. 5. Which capabilities can be distilled into reliable SLMs without unacceptable loss. 6. Whether edge inference will become close enough to free for widespread institutional deployment. 7. Whether local weights will meaningfully reduce dependence once hardware, data, updates, certification, and escalation are included. 8. Whether AI will materially lower the absorptive-capacity requirements for its own diffusion. 9. Whether cross-border AI access will reduce international inequality in welfare while increasing inequality in productive capacity or power. ## 18.3 Core synthesis The existing literature supports a more precise statement than “Western consumers transfer surplus to users in developing countries”: > Globally integrated platforms can use revenue concentrated in high-value markets to finance shared, low-marginal-cost systems that create substantial welfare in lower-monetization markets. This may reduce inequality in consumption welfare while leaving—or increasing—inequality in profits, productive capacity, institutional control, and political power. AI adds a potential discontinuity: > If useful model capability can be copied, legally transferred, locally operated, adapted, and maintained at low cost, internationally financed innovation may diffuse not merely as a service but as a productive asset. Whether that possibility becomes reality depends on technical reliability, excludability, complementary infrastructure, and systemic diffusion. # Part III — AI adoption, financing, and diffusion paths --- # 19. The variables that determine AI’s distributive form The distributional consequences of AI will not follow automatically from model capability. They will depend on choices and constraints across at least nine dimensions. | Dimension | Central question | |---|---| | Development finance | Who pays to train frontier and specialist models? | | Inference finance | Who pays each time a model is used? | | Deployment location | Does inference occur in a foreign cloud, regional infrastructure, or locally controlled hardware? | | Excludability | Can a provider restrict, meter, revoke, or geographically limit access? | | Ownership | Who possesses the weights, software, data, and hardware? | | Adaptability | Who may fine-tune, distill, connect, inspect, or modify the system? | | Complements | Who controls live data, sensors, identity, payments, distribution, and human escalation? | | Validation | Who determines that the system is accurate and safe enough for a particular setting? | | Diffusion capacity | Can local organizations redesign workflows and compound the resulting capability? | Different combinations produce different equilibria. “AI access” is therefore not one institutional arrangement but a family of arrangements ranging from revocable access to a foreign API to complete local possession of an open, validated system. --- # 20. Path A: rich-market enterprise demand finances a global free tier ## 20.1 Mechanism The most direct continuation of the Meta pattern is a freemium global AI platform. Enterprises and affluent professionals pay for high-value capabilities, while consumers elsewhere receive free or inexpensive access. Paying uses may include: - software development; - legal and financial analysis; - customer-service automation; - sales and marketing systems; - enterprise search; - scientific research; - workflow agents; - premium personal assistants; - dedicated capacity and contractual guarantees. A simplified chain is: ```text High-wage labor and high-value business processes → strong willingness to pay for automation or augmentation → enterprise AI revenue → shared model and infrastructure costs → free or cheap global consumer access ``` The commercial value of replacing or augmenting a worker is partly related to the worker’s wage and the value of the affected process. This initially favors monetization in rich economies even when usage spreads globally. ## 20.2 Why the free tier may be sustainable A provider may rationally serve low-revenue users because they: - impose low marginal costs as inference becomes more efficient; - support brand and distribution; - generate product feedback and safety information; - encourage developers to adopt the provider’s ecosystem; - create future premium users and enterprise customers; - forestall competitors; - strengthen a claim to global legitimacy; - help establish technical and interaction standards. A free tier may therefore be contribution-positive, strategically valuable, or both. Its existence does not prove an operating subsidy. ## 20.3 Likely quality stratification Unlike messaging, AI inference consumes material compute for each use. Free access may consequently differ from paid access in: - model capability; - number of queries; - context length; - response latency; - access to tools and live data; - reliability and uptime; - privacy protections; - agentic permissions; - customer support; - human escalation; - contractual warranties and liability. The likely result is not a simple binary divide between AI haves and have-nots. It is a continuous hierarchy of intelligence quality and institutional guarantees. ## 20.4 Distributional assessment **Potential benefits** - rapid global access; - low user price; - continuing model improvement; - reduced need for local technical operation; - broad language and accessibility interfaces. **Potential costs** - continuous foreign dependency; - discretionary service quality; - inability to audit or modify core systems; - exposure to pricing and policy changes; - unequal capability hidden beneath a universal interface; - concentration of usage data and organizational learning. **Classification:** highly plausible commercial path; already visible in current AI markets. --- # 21. Path B: advertising, recommendations, and agent-mediated commerce ## 21.1 From attention markets to decision markets An AI assistant can be offered without a monetary price if commercial actors pay for access to user attention, intent, recommendations, or completed transactions. Possible sources of finance include: - conventional advertisements; - sponsored answers or products; - referral and lead-generation fees; - commissions on purchases; - payments from lenders, insurers, healthcare providers, travel companies, or merchants; - preferential placement in agent search; - fees for access to transaction or identity rails. This architecture can go beyond advertising. Search engines monetize expressed intent; an agent may interpret, shape, and execute that intent. ## 21.2 Principal–agent conflict Consider a no-cost agricultural assistant financed by seed companies, fertilizer suppliers, insurers, commodity buyers, or lenders. The assistant may deliver real agronomic value while also influencing purchases and contractual relationships. The central question is: > Is the system acting as the user’s agent, the funder’s distribution channel, or both? Relevant safeguards would include: - clear disclosure of commercial relationships; - separation of advice and sponsorship; - comparison among providers; - explanation of recommendation criteria; - fiduciary or best-interest duties in sensitive domains; - auditable transaction logs; - meaningful refusal and human-review options. ## 21.3 Geographic effect Commercial intermediation could finance users in low-income markets without relying entirely on Western enterprise subscriptions. As local commerce digitizes, local merchants and financial institutions may increasingly fund the assistant. This reduces the cross-border subsidy while potentially increasing platform control over domestic transactions. A service can become more locally monetized and less locally sovereign at the same time. ## 21.4 Distributional assessment **Potential benefits:** free access, integrated action, product discovery, lower search costs, and viable local business models. **Potential costs:** concealed persuasion, discriminatory offers, conflicts of interest, extraction of transaction rents, and dependence on the agent as a commercial gatekeeper. **Classification:** plausible and potentially powerful; exact institutional forms remain unsettled. --- # 22. Path C: philanthropy and development finance ## 22.1 The rationale Some AI applications may produce large social returns but little private revenue. Examples include: - low-resource-language systems; - agricultural advice for subsistence farmers; - clinical support in underfunded health systems; - public-health surveillance; - special-needs education; - public-benefit navigation; - tools for small local governments; - datasets and evaluations for neglected populations. Foundations, aid agencies, development banks, and international organizations can finance the gap between social value and commercial return. ## 22.2 Four things a subsidy can purchase A subsidy can fund fundamentally different assets: 1. **Usage:** API credits or paid queries. 2. **Access infrastructure:** devices, electricity, connectivity, and cloud contracts. 3. **Deployment capability:** integration, training, data work, evaluation, and organizational redesign. 4. **Transferable capacity:** weights, open tools, documentation, local compute, maintainers, and the legal rights required for independent operation. The first is easiest to announce and the least durable. The fourth is hardest but most likely to create local capability. ## 22.3 Pilot dependency A common failure mode is a donor-funded pilot that succeeds technically but lacks a durable operating model. Risks include: - inference costs after grant expiry; - unavailable maintainers; - provider price or policy changes; - incompatibility with government procurement; - poor integration with public systems; - no institutional owner; - insufficient funds for evaluation and security updates. The relevant sustainability test is: > Can the recipient institution continue, modify, or replace the service after the grant and original vendor disappear? ## 22.4 Better philanthropic leverage Where feasible, philanthropy may create more durable public value by financing: - open evaluation suites; - local-language and domain datasets with legitimate governance; - reference implementations; - independent certification institutions; - model and adapter registries; - local technical training; - shared regional compute; - interoperability standards; - procurement and liability templates; - open or transferable specialist models. This shifts philanthropy from repeatedly purchasing consumption to lowering the fixed cost of local production and adoption. **Classification:** established financing path; long-run institutional effects depend on subsidy design. --- # 23. Path D: public universal service ## 23.1 AI as a publicly guaranteed service Governments could define a basic level of AI access as public infrastructure, analogous in selected respects to education, libraries, postal service, telephone access, or digital identity. Possible mechanisms include: - public procurement of citizen access; - vouchers or compute credits; - AI access through schools, libraries, clinics, and local government; - regulated universal-service obligations; - levies on commercial AI use; - publicly funded models and inference infrastructure; - regional purchasing consortia; - mandatory low-cost tiers for designated uses. ## 23.2 What a universal-service obligation must specify A meaningful guarantee requires more than nominal access. It must define: - eligible users and institutions; - covered tasks; - minimum model quality; - supported languages and modalities; - privacy and data-use rules; - uptime and continuity; - usage allowances; - accessibility requirements; - appeal and redress; - human escalation; - funding source; - audit and procurement requirements. Without minimum quality and continuity, “universal AI” could mean universal access to a materially inferior tier. ## 23.3 Financing choices Potential tax or levy bases include: - provider profits; - capital income; - large-scale compute or energy use; - tokens or inference volume; - automation-related savings; - commercial agent transactions; - general taxation. A token tax is intuitively connected to AI use but economically awkward. Tokens are not comparable across models or tasks; efficient models use fewer tokens; inference can move across jurisdictions; and beneficial low-margin uses could be discouraged. Taxes on rents, profit, transactions, or broad capital income may be less technologically precise but more administrable. **Classification:** normative institutional possibility; implementation would vary sharply by jurisdiction. --- # 24. Path E: geopolitical subsidy and AI soft power ## 24.1 AI provision as statecraft States may subsidize models, compute, cloud access, connectivity, and training abroad for strategic reasons. Historical analogues include satellite navigation, overseas broadcasting, development finance, telecommunications equipment, technical standards, and educational institutions. A subsidized model can spread: - language and cultural assumptions; - technical standards; - safety and censorship rules; - developer ecosystems; - cloud dependence; - commercial complements; - diplomatic influence. ## 24.2 Plausible blocs Possible strategies include: - US-supported access to commercial and open models for allies; - Chinese models bundled with cloud, telecom, device, and infrastructure finance; - European provision centered on regulatory compliance and public-sector use; - Indian support for Indic-language models and digital-public-infrastructure integration; - Gulf-funded Arabic-language systems; - regional African, Latin American, or Southeast Asian compute and model consortia. ## 24.3 Bargaining by recipient states Recipient governments may seek: - local data storage; - domestic inference or data centers; - low-cost public-sector access; - local-language support; - technology transfer; - training and employment; - tax revenue; - audit rights; - continuity guarantees; - routing to domestic model providers; - locally governed data and evaluation. This resembles Iridium’s political economy: a technology that can technically cross borders is re-territorialized through licenses, local partners, taxes, localization, and negotiated rents. ## 24.4 Managed interdependence Full self-sufficiency is unrealistic for most countries because chips, fabrication, cloud systems, energy, software, models, and talent are globally distributed. Practical sovereignty may instead mean: - preserving choice among foreign and domestic suppliers; - localizing sensitive components; - maintaining fallback capacity; - retaining authority over public-sector deployment; - avoiding irreversible dependence on one provider or state. The World Bank’s 2025 framework identifies connectivity, compute, context, and competency as the four foundations of inclusive AI participation, while noting that high-income countries continue to dominate innovation, compute infrastructure, and startup finance. **Classification:** already emerging; long-term bloc structure remains uncertain. --- # 25. Path F: sovereign and open AI ecosystems ## 25.1 What “sovereign AI” can mean The phrase is used for several different ambitions: 1. complete national control of the technical stack; 2. domestic frontier-model training; 3. local control of sensitive inference and data; 4. language and cultural representation; 5. policy autonomy and bargaining power; 6. the ability to switch suppliers; 7. continuity under sanctions, commercial withdrawal, or geopolitical crisis. These goals should not be conflated. A country can obtain meaningful operational autonomy without manufacturing frontier chips or training a leading general model. ## 25.2 Open models as leverage Open weights may enable: - local inference; - adaptation to language and domain; - inspection and independent evaluation; - competitive hosting; - continued operation after vendor exit; - training of local engineers; - creation of domestic complements. But “open” is not binary. Relevant questions include: - Are weights downloadable? - Are commercial use and redistribution allowed? - Is distillation allowed? - Are training data and recipes documented? - Are safety and evaluation tools available? - Can the model run on affordable hardware? - Are important dependencies proprietary? Open weights without affordable compute or local competence may yield little practical sovereignty. ## 25.3 Regional rather than national infrastructure For many lower-income countries, regional cooperation may be more economical than a national frontier effort. Shared institutions could provide: - compute clusters; - language and domain datasets; - security and evaluation laboratories; - public procurement; - legal and technical standards; - model registries; - incident reporting; - specialist training. This can pool fixed costs while preserving more public control than dependence on one foreign provider. **Classification:** plausible and partly observable; effectiveness depends on governance and sustained operating capacity. --- # 26. The technical fork: cloud service or local capital good ## 26.1 Cloud architecture ```text User or institution → provider-controlled API → provider-controlled model and compute → response or action ``` Advantages include frontier capability, centralized maintenance, rapid updates, and elastic capacity. Dependencies include connectivity, payment, authentication, provider policy, data transfer, and continuing service. ## 26.2 Local architecture ```text Local user data + local knowledge + selected live feeds → locally controlled model and tools → response or action ``` Advantages can include offline operation, low marginal financial cost, privacy, latency, continuity, and local policy control. Constraints include device memory, energy, model capability, maintenance, security, and evaluation. ## 26.3 Hybrid architecture ```text Request → local router or confidence gate → local SLM for routine cases → regional specialist model for bounded difficult cases → frontier cloud model for exceptional cases → human expert when risk or uncertainty requires it ``` This is a likely equilibrium because it allocates expensive capability selectively. Research on model cascades and routers already shows that systems can reduce cost by sending only some requests to more capable models. FrugalGPT reported matching its strongest individual model with up to a 98% cost reduction in its experimental settings. RouteLLM reported reductions of more than twofold in certain evaluations without sacrificing response quality. These are benchmark results, not guarantees for high-stakes deployment. ## 26.4 The political meaning of routing Routing is not merely a technical optimization. It decides: - which users receive which quality of model; - when a local system is considered inadequate; - which provider receives revenue and data; - when sensitive information leaves a jurisdiction; - whose risk threshold governs escalation; - whether poorer users receive systematically weaker treatment. A routing layer can reduce dependence on one model while becoming a new point of control itself. --- # 27. Small language models: evidence and limits ## 27.1 Why SLMs matter An SLM can be attractive where a task is bounded, usage is frequent, connectivity is limited, privacy is important, or cloud costs are material. Potential applications include: - classification and extraction; - translation and transcription; - form completion; - equipment troubleshooting; - crop or animal-health support combined with vision models; - protocol-based clinical support; - tutoring within a defined curriculum; - public-service navigation; - retrieval over a controlled local corpus. ## 27.2 Evidence of technical progress An ACL 2025 study evaluated 68 publicly accessible SLMs ranging from 100 million to 5 billion parameters. It found rapid improvement across the model class and reported that leading SLMs could match or exceed older 7B–8B reference models on selected general benchmarks. The study also found limited in-context-learning ability among some SLMs and significant remaining opportunities for hardware and runtime optimization. The result supports practical viability, not equivalence with frontier systems. Benchmarks can be contaminated or narrow; average accuracy can conceal rare but consequential failures; and a model competitive with an older 7B system is not necessarily suitable for autonomous high-stakes use. ## 27.3 How models become smaller Relevant techniques include: - knowledge distillation; - supervised fine-tuning; - quantization; - pruning and sparsity; - low-rank adaptation; - vocabulary and context optimization; - hardware-aware architecture design; - retrieval over an external knowledge base; - tool use rather than internal memorization. Distillation can use teacher-generated labels, examples, critiques, preferences, reasoning demonstrations, or internal representations where the teacher is accessible. A smaller model may learn a bounded behavior without reproducing the teacher’s full capabilities. ## 27.4 “Complete on written knowledge” — four separate claims ### Corpus completeness Has the frontier model encountered most relevant digitized text? For mature, well-documented domains, marginal web text may add little. But much knowledge remains private, undigitized, local, oral, tacit, low-resource, or poorly represented. ### Propositional knowledge completeness Does the model contain the stable facts and procedures required for a bounded task? This may be plausible for established curricula, routine administrative processes, maintenance manuals, and mature clinical or agronomic guidance. ### Capability completeness Can the model reliably apply that knowledge to unusual, ambiguous, adversarial, or multimodal cases? This remains much less certain. Knowledge possession is distinct from calibrated reasoning and exception handling. ### System completeness Can the model’s answer alter the real outcome? Advice does not itself supply medicine, credit, seed, transport, legal authority, a functioning market, or a human specialist. ## 27.5 Stable model plus changing context A useful deployment can separate layers by update frequency: | Layer | Example contents | Typical update pattern | |---|---|---| | Base model | language, stable domain concepts, reasoning patterns | infrequent | | Domain adapter | crop, medical, legal, or curriculum specialization | occasional | | Local knowledge pack | local rules, language, formulary, available equipment | periodic | | Live feeds | weather, prices, outbreaks, inventory, schedules | continuous or frequent | | User context | symptoms, location, history, preferences | per interaction | | Policy and escalation | permitted actions, thresholds, responsible institution | when governance changes | This reduces the need to retrain a model whenever facts change. It also makes the sources of an answer more inspectable. ## 27.6 Number of models The ecosystem could contain many task, language, and regional variants, but maintaining thousands of separately trained models would create severe versioning and validation costs. A more likely pattern is: ```text compact base model + domain adapter + language or voice pack + local retrieval corpus + policy module + live-data connectors + evaluation profile = locally deployed system ``` The number of deployments may be enormous even if the number of base-model families remains modest. **Evidence status:** Strong evidence that local SLM deployment is technically viable for some tasks. Broad claims about near-free, reliable, high-stakes intelligence remain hypotheses. --- # 28. Training, reproduction, and inference economics ## 28.1 Three distinct costs Discussion often collapses: 1. **Frontier development cost:** research, data, experiments, and major training runs. 2. **Adaptation cost:** distillation, fine-tuning, retrieval construction, evaluation, and integration. 3. **Operating cost:** inference, devices, electricity, connectivity, maintenance, and support. These costs may be borne by different parties in different countries. ## 28.2 Near-zero reproduction is not zero deployment cost Once weights exist, copying them may be extremely cheap. But a working service still requires hardware and institutional complements. The relevant equation is: $$ C_{local\ system} = C_{device} + C_{energy} + C_{integration} + C_{data} + C_{evaluation} + C_{maintenance} + C_{support} $$ For frequent, bounded usage, these costs may still be far below recurring cloud charges or scarce human expertise. ## 28.3 Why local marginal cost matters Local inference can remove several frictions beyond token price: - foreign-currency payment; - metering and billing; - account and identity requirements; - network availability; - cloud latency; - provider rate limits; - external data transfer; - per-query authorization. A device already purchased for other purposes may make an additional query feel effectively free to the user, even though energy and hardware depreciation remain real. ## 28.4 The quality-adjusted cost curve The relevant comparison is not cost per token but cost per successfully completed task at an acceptable risk level: $$ QAC = \frac{C_{inference} + C_{review} + E[C_{error}]}{P(acceptable\ completion)} $$ where \(E[C_{error}]\) is the expected cost of failure. A cheap model with frequent consequential errors can be more expensive than a frontier model or human expert. **Classification:** conceptual framework. Reliable cross-market cost data remain limited and change rapidly. --- # 29. Validation may be scarcer than models ## 29.1 Why benchmark performance is insufficient A locally deployed model must work under actual conditions: - local dialects and code-switching; - low-quality microphones and cameras; - missing or contradictory records; - unusual disease or crop presentations; - unavailable recommended products; - local units and customs; - weak connectivity and power; - adversarial fraud; - different legal and institutional constraints. Success on a generic benchmark does not establish improved health, income, learning, or administrative outcomes. ## 29.2 Validation stack A credible system may need: 1. technical benchmark evaluation; 2. local-language and subgroup testing; 3. field trials against relevant outcomes; 4. calibration and abstention tests; 5. red-team and security evaluation; 6. workflow and human-factors testing; 7. post-deployment monitoring; 8. incident reporting and investigation; 9. signed model and data updates; 10. periodic recertification. ## 29.3 Institutional implication If models become abundant, trusted evaluators may become a central scarce complement. Universities, professional associations, regulators, standards bodies, public laboratories, and civil-society organizations may matter more than another model-training effort. This creates a strong case for funding evaluation as shared infrastructure rather than requiring each clinic, school, or municipality to validate independently. ## 29.4 Healthcare as a limiting case Research on healthcare SLMs identifies real opportunities from distillation, quantization, pruning, privacy-preserving local operation, and domain adaptation. It also identifies insufficient data, long-tail distributions, label imbalance, uneven benchmark performance, and the need for evaluation across confidence, safety, factuality, and the care continuum. No single model’s leadership across one medical benchmark should be treated as evidence of general clinical reliability. **Classification:** strong inference from deployment and safety literature. --- # 30. Model routers and the settlement layer for cognition ## 30.1 Intelligence as a heterogeneous market If no model is optimal across every combination of task, cost, speed, privacy, language, and risk, a routing layer can classify requests and select among providers. Stripe’s announced acquisition of OpenRouter is a clear commercial bet on this architecture. Stripe states that OpenRouter routes across more than 400 models from over 80 providers and can select according to task complexity, price, speed, and reliability. Stripe describes tokens as a central currency for AI businesses and positions the combined system as infrastructure for managing both revenue and compute expenditure. ## 30.2 Analogy to international telephone settlements A future AI task may combine: - a user-facing application; - a local SLM; - a frontier model; - a specialist model; - proprietary data; - external tools; - payment and identity; - human escalation. A settlement system can meter usage and allocate revenue among participants: ```text Completed task payment → application provider → model provider(s) → data and tool providers → compute and routing layer → human or institutional specialist ``` This resembles the economic coordination role of telecom settlement systems while operating at the level of cognitive tasks. ## 30.3 Decentralized supply, centralized coordination A multi-model router can reduce dependence on one model provider while increasing dependence on the router. It may influence: - model discovery; - traffic allocation; - provider revenue; - quality and latency tiers; - data flows; - pricing; - permissible uses; - subsidy eligibility. Supplier neutrality, decision neutrality, and governance neutrality are different properties. Offering many models establishes only the first. ## 30.4 Edge bypass versus edge integration Pure local inference bypasses remote routing and settlement. A hybrid system may preserve the router as the control plane: ```text local request → local policy and difficulty classifier → free local inference → paid regional inference → subsidized frontier escalation ``` A government or foundation could fund only high-value escalations rather than every query. This can make subsidies more targeted, but it also places consequential allocative power in routing rules. ## 30.5 Cognitive price discrimination A router can observe task value, urgency, geography, available budget, quality requirements, and payer identity. It could therefore assign different models, prices, latency, privacy, or escalation rights to different users. AI inequality may become difficult to observe because two users encounter the same interface while receiving materially different underlying capability. **Evidence status:** Multi-model routing is established and commercially significant. Its future role as a global settlement layer and instrument of price discrimination is a reasoned projection. ### Key sources - Stripe, “Stripe agrees to acquire OpenRouter to help businesses optimize token routing and usage” (2026). https://stripe.com/newsroom/news/stripe-agrees-to-acquire-openrouter - Lingjiao Chen, Matei Zaharia, and James Zou, “FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance” (2023). https://arxiv.org/abs/2305.05176 - Isaac Ong et al., “RouteLLM: Learning to Route LLMs with Preference Data” (2024; revised 2025). https://arxiv.org/abs/2406.18665 --- # 31. Labor-market transmission across borders ## 31.1 Consumer benefit and income loss can coexist A country may receive valuable free AI services while losing labor income from tasks automated in richer markets. Welfare transfer and production displacement can move in opposite directions. A fuller accounting is: $$ \Delta W_j = \Delta CS_j + \Delta productivity_j + \Delta wages_j + \Delta profits_j + \Delta tax_j - \Delta externalities_j $$ Positive consumer surplus does not ensure a positive net effect. ## 31.2 Offshore cognitive work India, the Philippines, and other service-exporting economies are exposed because they sell routine cognitive labor to rich-country firms. The relevant comparison for an American buyer may be: ```text AI system cost versus offshore worker cost ``` not AI system cost versus an American worker. Automation can therefore affect a lower-wage country even when domestic employers there have weak incentives to automate. Potentially exposed activities include: - customer support; - data processing; - software testing and routine development; - transcription and translation; - bookkeeping; - document review; - content operations; - standardized analytics. ## 31.3 Countervailing opportunities The same economies may develop comparative advantages in: - AI-enabled services; - model adaptation and evaluation; - local-language data and interfaces; - human-in-the-loop operations; - implementation for emerging markets; - process redesign; - specialist supervision; - export of low-cost AI deployment capability. India may be especially positioned to industrialize diffusion rather than dominate frontier training, given its technical workforce, software-services sector, domestic diversity, and digital public infrastructure. ## 31.4 Current evidence should be treated cautiously Recent US payroll evidence finds no widespread economy-wide displacement but a substantial and widening employment gap for workers aged 22–25 in highly AI-exposed occupations, driven primarily by reduced hiring and concentrated in more substitution-oriented jobs. The authors explicitly describe these findings as descriptive early indicators rather than causal estimates. Evidence on cross-border service displacement remains less mature. Forecasts by consultants, investors, and industry participants should not be treated as observed labor-market outcomes. **Classification:** plausible and important transmission path; magnitude and timing remain uncertain. --- # 32. Sector cases ## 32.1 Agriculture ### Potential architecture ```text compact multilingual model + crop and region adapter + image or sensor model + local agronomic corpus + weather, pest, price, and inventory feeds + escalation to extension worker ``` ### Potential benefit - wider access to agronomic guidance; - faster disease identification; - weather-responsive decisions; - lower extension cost; - translation of technical material; - improved connection to markets and services. ### Binding complements - trustworthy local data; - affordable devices and connectivity; - access to seed, fertilizer, credit, insurance, storage, and markets; - field validation; - accountability for bad recommendations; - independence from commercially conflicted sponsors. The model may solve an information bottleneck without solving the material bottlenecks that prevent action. ## 32.2 Healthcare ### Potential architecture - local patient-facing language interface; - protocol and formulary retrieval; - narrow vision or diagnostic models; - privacy-preserving local inference; - cloud or specialist escalation; - clinician authority over consequential decisions. ### Potential benefit - triage and navigation; - documentation; - decision support; - translation; - follow-up and adherence; - support where specialists are scarce. ### Binding complements - clinical validation; - privacy and consent; - local disease prevalence and treatment availability; - calibrated abstention; - professional liability; - integration into records and referral networks; - human capacity for escalated cases. Healthcare illustrates why near-zero inference cost does not imply low total cost of safe service. ## 32.3 Education ### Potential benefit - individualized explanation; - translation; - teacher support; - curriculum adaptation; - accessibility for disabilities; - low-cost practice and feedback. ### Risks and complements - answer substitution for learning; - reduced productive struggle; - inaccurate or culturally inappropriate material; - surveillance of minors; - weak teacher preparation; - unequal access to high-quality models; - dependency on foreign curricula or values. The relevant outcome is learning, not answer generation or usage. ## 32.4 Government services ### Potential benefit - form completion; - eligibility navigation; - multilingual citizen support; - case summarization; - administrative productivity; - easier access to public information. ### Risks and complements - due-process violations; - opaque eligibility decisions; - exclusion caused by model error; - insecure identity and personal data; - vendor lock-in; - lack of appeal; - poor underlying administrative records. The strongest uses may assist citizens and public servants while leaving legally consequential decisions under accountable human authority. --- # 33. Alternative equilibria ## 33.1 Centralized cognitive utility A small number of frontier providers supply most high-quality intelligence through APIs and consumer products. Free tiers create broad access; enterprise and wealthy-market revenue fund the system. **Welfare:** high access and consumer surplus. **Sovereignty:** low. **Diffusion:** rapid consumption, uneven productive adaptation. **Control point:** provider cloud. ## 33.2 Federated but centrally settled intelligence Many models compete, but one or a few routing, identity, cloud, and payment layers mediate access. **Welfare:** competition can lower cost and improve matching. **Sovereignty:** greater model choice, continued intermediary dependence. **Diffusion:** broad developer access. **Control point:** router and settlement layer. ## 33.3 Hybrid local–cloud intelligence Routine work runs on local or regional models; difficult cases escalate to frontier systems or humans. **Welfare:** potentially high with lower recurring cost. **Sovereignty:** partial and task-dependent. **Diffusion:** requires local maintenance and evaluation institutions. **Control point:** escalation criteria, specialist models, live data, and hardware. ## 33.4 Cognitive commons Open models, adapters, data packs, evaluations, and interoperability standards support locally controlled deployment. Public and philanthropic institutions fund neglected applications and shared complements. **Welfare:** potentially broad and durable. **Sovereignty:** comparatively high. **Diffusion:** depends heavily on local capacity. **Control point:** distributed across maintainers, certifiers, hardware suppliers, and institutions. ## 33.5 Fragmented sovereign blocs States or regional blocs require domestic models, data localization, restricted cross-border inference, and national compute. **Welfare:** potentially lower model quality or higher cost in smaller markets. **Sovereignty:** higher in some layers; dependence may shift to domestic incumbents. **Diffusion:** shaped by industrial policy. **Control point:** state and selected national providers. ## 33.6 Layered inequality Basic assistants become universal, while affluent institutions receive superior reasoning, tools, agents, data, privacy, guarantees, and human escalation. **Welfare:** material gains at every tier. **Sovereignty:** varies independently of access. **Diffusion:** widespread but unequal in quality and organizational impact. **Control point:** purchasing power and institutional capability. These equilibria are not mutually exclusive. Different sectors and countries may occupy different arrangements simultaneously. --- # 34. Policy and subsidy design principles ## 34.1 Purchase capability, not only consumption Where technically feasible, public and philanthropic funding should prefer assets that recipients can continue using, adapting, and replacing after external support ends. ## 34.2 Fund complements explicitly Model access without connectivity, context, competency, evaluation, and implementation is unlikely to generate systemic value. The World Bank’s four-Cs framework—connectivity, compute, context, and competency—is a useful minimum checklist. ## 34.3 Separate advice from commercial influence Where assistants are advertiser-, vendor-, or transaction-funded, require disclosure, comparison, auditability, and stronger duties in health, finance, education, agriculture, and public services. ## 34.4 Define minimum service quality Universal access programs should specify capability, language, privacy, continuity, escalation, and redress rather than counting registered accounts or nominal availability. ## 34.5 Build independent validation institutions Support local and regional bodies that can test systems under relevant languages, risks, workflows, and material conditions. ## 34.6 Preserve substitutability Use open standards, exportable data, model and provider portability, modular architectures, and public procurement terms that make replacement credible. ## 34.7 Avoid false sovereignty Domestic branding, a national data center, or a downloadable model may conceal dependence on foreign chips, software, updates, evaluation, and cloud escalation. Map dependencies by layer. ## 34.8 Measure distribution beyond access Track: - effective use and outcome improvement; - capability and quality by income and geography; - local wages, firms, taxes, and technical learning; - provider concentration; - institutional switching capacity; - externalities and incidents; - the durability of service after subsidies end. --- # 35. Testable hypotheses and indicators | Hypothesis | Supporting indicators | Disconfirming indicators | |---|---|---| | Rich-market enterprise revenue will finance global consumer AI access | falling free-tier prices; broad low-income-country availability; enterprise revenue concentration | widespread geographic withdrawal or strict cost-reflective pricing | | SLMs will make bounded intelligence locally operable | comparable task outcomes on affordable hardware; rising offline use; open transferable deployments | persistent large quality gaps; unacceptable error rates; hardware costs remain prohibitive | | Hybrid routing will dominate pure local or pure cloud architectures | growing confidence-gated escalation; standards for local/cloud orchestration; selective frontier spending | most tasks remain entirely centralized or entirely local | | Validation will become a primary bottleneck | model abundance with slow certified deployment; demand for local test institutions | reliable deployment follows model release with little evaluation burden | | AI will partially accelerate its own diffusion | novice implementers successfully integrating systems; reduced training and coding requirements | organizational and technical adoption costs remain unchanged | | Access can rise while sovereignty falls | increasing use alongside provider concentration and switching difficulty | broad local ownership, interoperable systems, and credible exit options | | Countries with stronger diffusion capacity will capture more productivity value from the same model access | divergent outcomes correlated with skills, institutions, data, and infrastructure | similar gains across countries regardless of complements | | Offshore cognitive-service economies face early transmission effects | reduced entry-level service hiring; contract compression; AI-enabled consolidation | stable employment and rising service exports despite client automation | | Subsidies that transfer assets outperform API-credit programs over time | lower recurring costs, local maintenance, provider substitutability | abandoned local systems and superior sustainability from managed cloud services | --- # 36. Interim conclusion AI can diffuse internationally through at least six financing paths: rich-market enterprise cross-subsidy, advertising and commerce, philanthropy, public universal service, geopolitical subsidy, and sovereign or open ecosystems. These mechanisms can coexist within the same product. The central technical fork is not simply large versus small models. It is whether useful capability remains tied to a provider-controlled service or can become a legally and practically transferable local asset. The likely architecture is hybrid. Routine tasks will increasingly run on smaller local or regional systems; difficult, current, or high-risk cases will escalate to frontier models and human experts. This can lower cost and increase autonomy without eliminating dependence. Control may migrate from the model provider to hardware suppliers, data owners, certifiers, routers, payment systems, or escalation services. The distributional outcome must therefore be evaluated at multiple levels: 1. Who receives useful service? 2. Who pays for it? 3. Who owns and controls the deployed capability? 4. Who captures productivity, profit, wages, taxes, data, and learning? 5. Who can continue, alter, or replace the system? 6. Which societies can diffuse it widely enough to compound the gains? Broad access can coexist with concentrated power. Local possession can coexist with weak diffusion. And foreign financing can create both genuine welfare and durable dependency. The strongest version of the optimistic case is not that every person receives free access to a frontier chatbot. It is that frontier research financed where willingness to pay is highest helps create validated, adaptable, locally governable systems that accumulate capability where income is lower. That remains possible. It is not the default outcome of technical progress alone. # Part IV — Argument audit, counterarguments, and sources --- # 37. Strongest counterarguments ## 37.1 There may be no meaningful cross-border subsidy The central framing may mistake consolidated corporate accounting for economic redistribution. A true operating subsidy requires evidence that revenue attributable to a market is below the avoidable cost of serving it. Public information does not establish that India or Brazil is contribution-negative for Meta. Software has large shared fixed costs and low marginal costs. Once the global platform exists, serving an additional user may be profitable even at very low revenue per user. Low-ARPU markets may also create substantial strategic value through network effects, international connections, product learning, competitive foreclosure, and future monetization. The relationship may therefore be reciprocal rather than redistributive. ### Best response Use “cross-border financing of locally created consumer surplus” as the broad claim. Reserve “operating subsidy” for cases with cost and counterfactual evidence. The economically interesting fact does not disappear: the geography of revenue, benefit, ownership, and control remains sharply separated. ### What would resolve it - country-level attributable revenue; - avoidable infrastructure, moderation, support, legal, and product costs; - internal investment and shutdown counterfactuals; - evidence about the strategic value of users in low-monetization markets. ## 37.2 Affluent consumers do not necessarily pay for WhatsApp The intuitive chain from consumer spending to advertiser revenue to Meta may overstate consumer incidence. Advertising may be financed through lower producer margins, shareholder returns, wages, or supplier payments. It can also improve matching and competition, potentially lowering prices rather than raising them. ### Best response The essay says Western consumer spending *indirectly finances* the system, not that every advertising dollar is passed through into consumer prices. Advertisers’ willingness to pay ultimately depends on expected commercial value from consumer behavior. A full incidence claim would require market-specific evidence. ## 37.3 Consumer-surplus estimates for free digital goods may be unreliable Willingness-to-accept estimates can be sensitive to elicitation design, time horizon, strategic responses, hypothetical bias, income, and the absence of substitutes. High valuations may measure network dependence or withdrawal costs rather than improvements in well-being. Moreover, private willingness to accept does not automatically include harms to non-users, privacy loss, misinformation, fraud, political externalities, or displacement of competing services. ### Best response Treat these estimates as evidence that zero-price digital goods create large unpriced benefits, not as exact welfare accounts or additions to GDP. The principal argument requires only that WhatsApp is highly valuable to many users, not a precise dollar valuation. The underlying experimental method has been published in peer-reviewed work using incentivized choices, which strengthens it relative to an ordinary opinion survey while leaving the conceptual limitations intact. ## 37.4 “Consumer surplus is not sovereignty” may combine unlike concepts Consumer surplus is an individual welfare concept. Sovereignty concerns authority, legitimacy, control, and collective self-determination. Their contrast may sound rhetorically sharp while comparing different analytical levels. ### Best response The distinction is deliberate. It warns against inferring governance legitimacy, national capability, or control from demonstrated user benefit. The two concepts should not be collapsed into one metric. A useful foreign service can improve individual welfare while increasing collective dependency. A further refinement is necessary: control is not by itself legitimate sovereignty. Private companies can exercise quasi-sovereign control without democratic or public legitimacy. State control can also be coercive or incompetent. Analysis should therefore ask both **who controls** and **whether that control is legitimate and accountable**. ## 37.5 Local execution may not create meaningful sovereignty An edge model can still depend on foreign chips, operating systems, frameworks, app stores, licenses, updates, data feeds, and cloud escalation. Local possession of weights may be a thin layer atop a deeply foreign stack. ### Best response Agreed. The companion treats sovereignty as layered and graded rather than binary. Local execution can improve continuity, privacy, bargaining power, and marginal cost without producing complete autonomy. The relevant goal for most countries is managed interdependence and credible substitution, not complete national self-sufficiency. Recent sovereignty analysis likewise maps dependencies across physical inputs, energy, chips, cloud, networks, frameworks, data, models, talent, and governance, concluding that a fully domestic stack is nearly impossible for most countries. It recommends selective control, diversification, open standards, and investment where local comparative advantage exists. ## 37.6 Small models may remain too unreliable for the proposed uses The most rhetorically compelling applications—medicine, agriculture, education, and public services—are context-dependent and consequential. Small models may perform well on selected benchmarks while failing on rare cases, local dialects, poor-quality inputs, adversarial prompts, and out-of-distribution conditions. ### Best response The essay presents local SLM deployment as a possibility, not an accomplished fact. Current evidence supports practical edge viability for some tasks, not autonomous reliability across high-stakes domains. The strongest early cases are likely to be bounded, frequent, verifiable, low-risk tasks with retrieval, tools, confidence gates, and human escalation. A major ACL study of 68 public SLMs found rapid capability gains and performance competitive with older 7B–8B models on selected tasks, while also finding limited in-context learning and substantial remaining efficiency constraints. This evidence supports the fork without resolving it. ## 37.7 The frontier may keep moving faster than edge capability Even if today’s frontier capability can be distilled, users may continually demand better reasoning, modalities, agency, current knowledge, and reliability. Local models may become the equivalent of old operating systems: still functional but increasingly inadequate. Dependence can persist through the growing quality gap rather than formal revocation. ### Best response This is a serious possibility and supports a hybrid rather than fully local base case. The key empirical question is whether bounded socially valuable tasks have a capability threshold after which further frontier gains produce diminishing practical returns. If they do, local systems can remain useful for long periods. If task expectations continually expand, frontier dependence will persist. ## 37.8 Distillation may be legally or technically restricted The local-capital thesis assumes frontier capabilities can be transferred into smaller models. Providers may prohibit output harvesting, restrict distillation contractually, limit rate access, watermark outputs, pursue patents, or keep essential training and evaluation methods secret. ### Best response The distinction between technical reproducibility and legal transferability is central. Authorized distillation is common, but unauthorized distillation occupies an unsettled legal space involving contract, computer-access law, copyright, trade secrets, and potentially patents. The optimistic thesis requires models or outputs that recipients are legally entitled to use, modify, and redistribute—not merely an ability to extract them covertly. ## 37.9 Model ownership may distract from better managed services A cloud provider may offer superior security, updates, monitoring, uptime, and economies of scale. Requiring every clinic, school, or municipality to operate local models could reproduce the failures of poorly maintained local IT systems. Sovereignty rhetoric may encourage expensive national prestige projects rather than reliable service. ### Best response Local ownership is not always preferable. The correct architecture depends on task risk, usage volume, connectivity, privacy, local competence, and switching options. Managed cloud services can be the better choice when procurement preserves portability and oversight. “Purchase capability, not only consumption” is a design preference where feasible, not a rule that every inference must occur locally. ## 37.10 Open models can concentrate power too Open weights may favor organizations with the compute, data, distribution, and talent to exploit them. They can also support surveillance, cybercrime, propaganda, or authoritarian control. A domestic model can empower a local state or incumbent rather than citizens. ### Best response Openness redistributes some technical options; it does not guarantee democratic outcomes. Sovereignty must include legitimacy, accountability, rights, and access to alternatives. The relevant comparison is among concrete governance arrangements, not “open” versus “closed” in isolation. ## 37.11 Diffusion may increase inequality within countries Even if lower-income countries gain, the benefits may accrue to urban firms, skilled workers, political elites, and well-connected institutions. Rural users may receive low-quality interfaces while domestic incumbents capture productivity gains and rents. ### Best response Cross-country convergence can coexist with within-country divergence. Distribution should be measured by income, region, language, gender, disability, firm size, and institutional quality—not only national averages. ## 37.12 AI could destroy more tradable income than consumer surplus creates India or the Philippines might receive useful free assistants while losing export jobs in business-process outsourcing, customer service, and routine software work. The welfare value of free cognition may be smaller than the lost wages, tax base, and skill-formation pathway. ### Best response This is an unresolved net-welfare question and one of the strongest objections to an equalization narrative. Analysis must include labor income, producer surplus, taxes, capability accumulation, and transition costs. Consumer surplus alone cannot settle it. ## 37.13 Universalist rhetoric may be causally irrelevant Companies may invoke remote farmers and underserved patients, but capital allocation is driven by enterprise revenue and strategic competition. The rhetoric may neither cause deployment nor meaningfully constrain it. ### Best response The claim is not that rhetoric determines economics. Rhetoric matters because it legitimizes investment, shapes regulation, attracts employees and partners, and frames corporate expansion as social progress. Its practical influence should be studied rather than assumed. --- # 38. Alternative explanations and competing frames ## 38.1 Ordinary global innovation spillover The simplest explanation is that AI follows a familiar pattern: rich markets finance expensive innovation; competition and low reproduction costs spread benefits internationally. No novel redistribution concept is required. **Usefulness:** highlights continuity with software, medicine, and scientific knowledge. **Blind spot:** underplays platform governance, cross-product financing, and infrastructure dependence. ## 38.2 Multisided exchange rather than redistribution Users, advertisers, merchants, developers, and institutions contribute different forms of value. Low-ARPU populations exchange network participation and future option value for free service. **Usefulness:** avoids treating users as passive aid recipients. **Blind spot:** may obscure extreme asymmetry in bargaining power and value capture. ## 38.3 Global public-good provision Frontier knowledge is expensive to create and cheap to reproduce. Rich-market financing supports a non-rival asset whose global diffusion is socially efficient. **Usefulness:** clarifies why zero or low marginal prices can maximize welfare. **Blind spot:** weights, compute, data, validation, and deployment are not pure public goods. ## 38.4 Digital dependency or colonialism Foreign platforms provide useful infrastructure while concentrating data, standards, profit, and control in the provider’s home jurisdiction. **Usefulness:** foregrounds power, exit, and productive capacity. **Blind spot:** can understate voluntary adoption, reciprocal value, and genuine consumer welfare. ## 38.5 Global value-chain specialization Countries need not own every layer. They may rationally specialize in deployment, adaptation, services, data, chips, energy, or applications while importing other layers. **Usefulness:** rejects autarkic definitions of sovereignty. **Blind spot:** market specialization can become dangerous dependence when suppliers are concentrated or geopolitically constrained. ## 38.6 Development aid embedded in commerce Commercial systems can generate large positive externalities outside the markets supplying most revenue. The result resembles aid without deliberate transfer or recipient governance. **Usefulness:** captures the unusual geography of benefit. **Blind spot:** “aid” can conceal the strategic value received by the provider and the absence of recipient ownership. No single frame should dominate. The object is best understood as a joint system of welfare creation, commercial exchange, innovation spillover, and asymmetric control. --- # 39. Failure modes for the optimistic AI thesis The proposition that AI could distribute locally owned productive intelligence would fail or be materially weakened under any combination of the following conditions. ## 39.1 Capability failure Small or local models remain substantially inferior on economically valuable tasks, and frontier capability cannot be compressed without unacceptable losses. ## 39.2 Reliability failure Average performance improves, but long-tail errors, weak calibration, and poor abstention make local systems unsafe without expensive human review. ## 39.3 Cost failure Hardware, energy, integration, maintenance, and evaluation costs remain too high for widespread deployment, even if weights are cheap. ## 39.4 Data failure Useful performance requires proprietary, current, or highly local data that recipients cannot obtain or govern. ## 39.5 Legal enclosure Licenses, contracts, patents, trade secrets, export controls, app-store rules, or technical access controls prevent meaningful local adaptation and redistribution. ## 39.6 Complement failure Weak connectivity, institutions, supply chains, capital, procurement, and technical skills prevent model possession from improving outcomes. ## 39.7 Validation failure No affordable trusted institutions emerge to evaluate and certify local systems. Model abundance produces uncertainty rather than adoption. ## 39.8 Update failure Threats, standards, knowledge, and user expectations change quickly enough that disconnected models become unsafe or obsolete. ## 39.9 Market-concentration failure Open or local models exist but distribution, hardware, identity, payments, routing, or certification consolidate around a few gatekeepers. ## 39.10 Political failure States use sovereignty policy to entrench domestic incumbents, expand surveillance, suppress information, or finance prestige models instead of broad capability. ## 39.11 Labor-transition failure AI eliminates tradable service jobs and entry-level learning pathways faster than recipient economies develop new capabilities and industries. ## 39.12 Legitimacy failure Systems become socially indispensable without accountable mechanisms for rule-setting, appeal, redress, and citizen participation. --- # 40. Claim-by-claim evidence audit | ID | Claim | Status | Confidence | Best current support | Principal gap | |---|---|---|---|---|---| | C1 | Iridium achieved broad technical coverage but was initially unaffordable for mass low-income use | Documented historical claim | High | Bloom; Iridium histories and pricing records | Exact relationship between public rhetoric and internal market expectations | | C2 | Meta’s monetization per user varies dramatically by geography | Documented at regional level; country figures partly proprietary or estimated | High regionally | Meta financial disclosures and informed internal estimates | Public country-level ARPU/ARPP and product allocation | | C3 | WhatsApp functions as quasi-public infrastructure in India and parts of South America | Literature-backed descriptive claim | High | Usage, business, public-service, and communications evidence | Consistent cross-country definition of “public utility” function | | C4 | Rich-market advertising revenue helps finance global WhatsApp operations | Strong inference from consolidated economics | High | Meta revenue composition and global operations | Product- and country-specific cost allocation | | C5 | India is a true operating loss for Meta | Unproven | Low from public data | Directional internal knowledge may support it | Avoidable-cost and counterfactual evidence | | C6 | Western consumers ultimately finance part of Meta’s global system | Qualified incidence inference | Medium–high | Advertiser demand depends on expected consumer value and spending | Incidence across prices, profits, wages, and suppliers | | C7 | Free digital goods create large consumer surplus | Empirically supported | High | Brynjolfsson et al.; PNAS choice experiments | Externalities and interpretation of WTA | | C8 | Lower-income countries receive proportionately larger digital-welfare gains | Empirically supported in studied sample | Medium–high | *Digital Welfare of Nations* | Generalization beyond sampled countries and goods | | C9 | Consumer surplus does not imply ownership or legitimate control | Conceptual distinction | High | Platform-governance and sovereignty literature | Operational measurement of dependency and legitimacy | | C10 | Innovators often capture a minority of total social value | Strong literature-backed claim | High | Arrow, Nelson, Nordhaus | AI-specific appropriability may differ sharply | | C11 | International knowledge diffusion can disproportionately benefit smaller or R&D-poor countries | Model- and literature-backed claim | Medium–high | Cai, Li, and Santacreu; diffusion literature | AI-specific empirical evidence | | C12 | Absorptive capacity determines whether external knowledge becomes productive | Established organizational theory | High | Cohen and Levinthal | Translation from firms to countries and AI-specific measurement | | C13 | Current SLMs are practically viable on edge devices for some tasks | Empirically supported | High for bounded tasks | Lu et al., ACL 2025 | Field reliability, newer devices, and high-stakes tasks | | C14 | SLMs can match larger models on selected benchmarks | Empirically supported | High with benchmark qualification | Lu et al. | Contamination, robustness, and real-world outcomes | | C15 | Frontier models are “complete” on relevant written knowledge | Contested hypothesis | Low–medium depending on domain | Saturation intuitions and broad pretraining coverage | Undigitized, tacit, local, private, and changing knowledge | | C16 | Distillation can convert frontier capability into local capital goods | Technically plausible | Medium | Distillation practice and edge models | Legal rights, capability retention, validation, and maintenance | | C17 | Local inference approaches zero marginal cost | Qualified hypothesis | Medium for some tasks | Existing on-device deployment and falling hardware cost | Quality-adjusted cost, energy, depreciation, support | | C18 | Local execution increases sovereignty | Conditional inference | Medium | Continuity, privacy, and provider-independence logic | Upstream chips, software, licenses, updates, and data | | C19 | Hybrid local/cloud routing can lower cost substantially | Empirically supported in selected settings | High for tested routing tasks | FrugalGPT; RouteLLM | High-stakes calibration and production generalization | | C20 | Routing layers may become settlement and control points | Forward-looking inference | Medium | OpenRouter/Stripe strategy and platform history | Future market structure and interoperability | | C21 | Validation may become scarcer than model supply | Hypothesis | Medium | Deployment literature and regulated-domain requirements | Comparative evidence across sectors and countries | | C22 | AI may accelerate its own diffusion | Hypothesis | Medium | Coding, translation, and instructional capability | Organization-level causal evidence | | C23 | Diffusion capacity matters more than frontier invention for national AI gains | Historically grounded hypothesis | Medium–high | Ding’s framework | AI may differ from earlier general-purpose technologies | | C24 | Offshore service economies may experience early job displacement | Plausible hypothesis | Medium | Task exposure and early rich-country employment evidence | Direct causal evidence in India, Philippines, and similar economies | | C25 | Asset-transferring subsidies outperform API credits | Normative empirical hypothesis | Medium–low | Sustainability and capability-building logic | Comparative longitudinal evaluations | | C26 | Universal AI access will conceal substantial quality stratification | Forecast | Medium–high | Existing free/premium tiers and inference economics | Future competition, regulation, and cost curves | | C27 | AI could distribute productive capacity more widely than prior global platforms | Central hypothesis | Medium | Reproducibility, open weights, edge deployment | All capability, cost, legal, and diffusion conditions jointly holding | --- # 41. Research agenda ## 41.1 Platform financing and geography 1. Estimate avoidable cost and attributable revenue by geography for major global digital services. 2. Separate product-level and country-level cross-subsidies inside conglomerate platforms. 3. Measure the strategic and network value contributed by low-ARPU users. 4. Study who bears advertising expenditure across industries and market structures. 5. Develop welfare measures that combine consumer surplus with privacy, misinformation, fraud, competition, and political externalities. ## 41.2 AI deployment economics 1. Publish quality-adjusted cost curves for local, regional, and frontier-cloud models by task. 2. Include hardware depreciation, energy, integration, review, error, and maintenance—not only token price. 3. Identify tasks with stable performance thresholds beyond which frontier gains have low marginal social value. 4. Measure the operating durability of local models under changing knowledge and security conditions. ## 41.3 Sovereignty and substitutability 1. Build layer-by-layer dependency maps for public-sector AI deployments. 2. Measure switching cost, data portability, model portability, and continuity under provider exit. 3. Compare open-weight, managed-cloud, and hybrid systems on actual institutional autonomy. 4. Distinguish effective control from legitimate authority and public accountability. ## 41.4 Diffusion and capacity 1. Compare outcomes from identical or similar models deployed in countries with different absorptive capacity. 2. Measure whether AI tools enable novice implementers to perform integration work previously requiring advanced specialists. 3. Study whether local model deployment creates durable firms, jobs, tax revenue, and technical learning. 4. Track diffusion through ordinary firms and public agencies rather than benchmark leaders and national flagship models. ## 41.5 Subsidy design 1. Compare API credits, managed-service procurement, local infrastructure, open-model transfer, and evaluation funding. 2. Test sustainability after donor or pilot funding ends. 3. Measure recipient control, provider substitutability, and local capacity accumulation. 4. Study when local operation is genuinely superior to well-governed shared cloud infrastructure. ## 41.6 Labor and distribution 1. Track entry-level and mid-level employment in service-exporting economies. 2. Distinguish augmentation from substitution by task and organizational design. 3. Measure whether free consumer AI compensates households for labor-income losses. 4. Study new export opportunities in adaptation, evaluation, human escalation, and deployment. 5. Examine distribution within countries by region, language, gender, education, disability, and firm size. --- # 42. Questions an AI assistant should ask the reader Rather than giving a generic summary, an assistant using this companion should ask which issue the reader wants to examine. Useful branches include: 1. **Economics:** Is the Meta/WhatsApp arrangement truly a subsidy? 2. **Incidence:** Who ultimately pays for advertiser-funded infrastructure? 3. **History:** What are the closest analogues in telecom, broadcasting, universal service, and global public goods? 4. **Measurement:** How should consumer surplus from a zero-price service be estimated? 5. **Sovereignty:** What would meaningful control over a local AI system require? 6. **Technology:** Which tasks are plausible for SLMs and edge deployment? 7. **Diffusion:** How does Jeffrey Ding’s framework change the policy prescription? 8. **Development:** Could high consumer welfare coexist with weak domestic value capture? 9. **Labor:** Could AI transfer consumption benefits while destroying tradable service income? 10. **Governance:** Who should set routing, quality, commercial-influence, and escalation rules? 11. **Policy:** Should subsidies fund queries, infrastructure, validation, or transferable assets? 12. **Falsification:** What evidence would show that the essay’s central thesis is wrong? --- # 43. Suggested prompts for readers ## 43.1 General exploration ```text I have read the essay “Who Pays for Intelligence?” Use its research companion as context, but do not assume its thesis is correct. Ask me which branch of the argument I want to explore. Distinguish documented evidence, inference, and speculation; present the strongest counterarguments; and cite original sources where possible. ``` ## 43.2 Adversarial critique ```text Critique the thesis of “Who Pays for Intelligence?” as a skeptical economist. Focus on the absence of a demonstrated counterfactual subsidy, advertising incidence, shared fixed costs, network effects, and problems with measuring consumer surplus from zero-price goods. Identify which claims survive the critique. ``` ## 43.3 Development-policy analysis ```text Apply the framework in “Who Pays for Intelligence?” to a specific country and sector. Evaluate access, consumer surplus, local value capture, sovereignty, absorptive capacity, labor effects, and institutional complements. Compare cloud, local, and hybrid deployment architectures. Do not assume that local operation is always preferable. ``` ## 43.4 Technical deployment analysis ```text Choose one bounded application proposed in the companion. Design a realistic local or hybrid AI architecture, including base model, adapters, retrieval, live feeds, hardware, routing, human escalation, validation, maintenance, and total cost. Identify which parts can be locally controlled and which remain externally dependent. ``` ## 43.5 Subsidy-design analysis ```text Compare five ways to subsidize this AI application: API credits, managed cloud procurement, local model transfer, regional shared infrastructure, and open evaluation capacity. Score each on cost, reliability, speed, local capability accumulation, substitutability, sovereignty, and post-grant sustainability. ``` ## 43.6 Research update ```text Update the companion’s factual claims using current primary sources and peer-reviewed research. Focus on SLM capability, edge inference costs, licensing, routing markets, labor effects, and cross-country AI adoption. Preserve the original claim IDs and show whether confidence should rise, fall, or remain unchanged. ``` --- # 44. Annotated bibliography ## 44.1 Platforms and cross-subsidy ### Rochet, Jean-Charles, and Jean Tirole. “Platform Competition in Two-Sided Markets.” *Journal of the European Economic Association* 1, no. 4 (2003): 990–1029. https://doi.org/10.1162/154247603322493212 Foundational model of competition and price allocation across interdependent platform sides. Supports the analysis of zero or below-cost pricing to users when their participation creates value for advertisers or other participants. Does not itself establish geographic subsidy. ### Armstrong, Mark. “Competition in Two-Sided Markets.” *RAND Journal of Economics* 37, no. 3 (2006): 668–691. https://doi.org/10.1111/j.1756-2171.2006.tb00037.x Develops models of platform competition, cross-group externalities, and single- versus multi-homing. Useful for understanding why model and router interoperability may affect concentration. ## 44.2 Digital consumer welfare ### Brynjolfsson, Erik, Avinash Collis, W. Erwin Diewert, Felix Eggers, and Kevin J. Fox. “GDP-B: Accounting for the Value of New and Free Goods in the Digital Economy.” NBER Working Paper 25695 (2019). https://doi.org/10.3386/w25695 Develops an approach to supplement conventional national accounts with welfare gains from free digital goods. ### Brynjolfsson, Erik, Avinash Collis, and Felix Eggers. “Using Massive Online Choice Experiments to Measure Changes in Well-Being.” *Proceedings of the National Academy of Sciences* 116, no. 15 (2019): 7250–7255. https://doi.org/10.1073/pnas.1815663116 Peer-reviewed presentation of incentivized single-binary-discrete-choice experiments used to estimate willingness to accept for digital goods. Establishes methodological seriousness but does not eliminate questions about network dependence, externalities, or interpretation. ### Brynjolfsson, Erik, Avinash Collis, Daley Kutzman, Harshita Liaqat, and coauthors. “The Digital Welfare of Nations: New Measures of Welfare Gains and Inequality.” NBER Working Paper 31670 (2023). https://doi.org/10.3386/w31670 Cross-country study of nearly 40,000 participants across 13 countries. Estimates large welfare gains from ten digital goods and larger proportional gains in lower-income countries. Highly relevant to the geography of digital welfare. Review disclosed author affiliations and financial relationships when interpreting the work. ## 44.3 Innovation, spillovers, and absorptive capacity ### Arrow, Kenneth J. “Economic Welfare and the Allocation of Resources for Invention.” In *The Rate and Direction of Inventive Activity* (1962). Canonical account of information, appropriability, uncertainty, and underinvestment in invention. ### Nelson, Richard R. “The Simple Economics of Basic Scientific Research.” *Journal of Political Economy* 67, no. 3 (1959): 297–306. Explains why firms may underinvest in basic research when benefits are difficult to appropriate. ### Nordhaus, William D. “Schumpeterian Profits in the American Economy: Theory and Measurement.” NBER Working Paper 10433 (2004). https://doi.org/10.3386/w10433 Estimates that producers captured only a small part of the social returns from technological advances in the studied US economy. The often-cited 2.2% figure should not be generalized mechanically to AI. ### Cohen, Wesley M., and Daniel A. Levinthal. “Absorptive Capacity: A New Perspective on Learning and Innovation.” *Administrative Science Quarterly* 35, no. 1 (1990): 128–152. https://doi.org/10.2307/2393553 Defines absorptive capacity as the ability to recognize the value of external information, assimilate it, and apply it commercially. Emphasizes prior related knowledge and path dependence. Essential for explaining why free access to an artifact does not guarantee productive adoption. ### Cai, Jie, Nan Li, and Ana Maria Santacreu. “Knowledge Diffusion, Trade and Innovation across Countries and Sectors.” Working paper (2017). https://economics.yale.edu/sites/default/files/cls-2017-santacreu.pdf Models international knowledge diffusion and finds that foreign knowledge spillovers amplify dynamic welfare gains, with smaller and R&D-poor countries potentially gaining more from external ideas. ### Ding, Jeffrey. *Technology and the Rise of Great Powers: How Diffusion Shapes Economic Competition*. Princeton University Press, 2024. https://press.princeton.edu/books/paperback/9780691260341/technology-and-the-rise-of-great-powers Argues that broad diffusion, institutional adaptation, and skill infrastructure matter more to national power during technological revolutions than isolated frontier breakthroughs. ## 44.4 Knowledge and global public goods ### Stiglitz, Joseph E. “Knowledge as a Global Public Good.” In *Global Public Goods*, 1999. Treats knowledge as non-rival and potentially global while explaining why production and dissemination require institutional finance. ### Reichman, Jerome H., and Keith E. Maskus. “The Globalization of Private Knowledge Goods and the Privatization of Global Public Goods.” *Journal of International Economic Law* 7: 279–320. https://scholarship.law.duke.edu/faculty_scholarship/2144/ Examines tensions between global intellectual-property protection, private knowledge goods, innovation incentives, and access to knowledge in developing countries. ## 44.5 Telecom, Iridium, and universal service ### Wallsten, Scott J. “Telecommunications Investment and Traffic in Developing Countries: The Effects of International Settlement Rate Reforms.” *Journal of Regulatory Economics* 20 (2001): 307–323. https://doi.org/10.1023/A:1011171110899 Finds no evidence that large settlement payments to developing-country carriers increased telephone penetration or equipment investment. Important warning that cross-border cash transfers may become incumbent rents rather than public benefit. ### International Telecommunication Union. “Accounting Rate Reform undertaken by ITU-T Study Group 3.” https://www.itu.int/en/ITU-T/studygroups/2013-2016/03/Pages/accounting-rate.aspx Primary institutional account of the international accounting-rate system and its reform. ### Bloom, John. *Eccentric Orbits: The Iridium Story*. Grove Press, 2016. Narrative history of Iridium’s engineering, financing, lobbying, bankruptcy, and rescue. Useful as a richly reported secondary source. Specific claims about agreements or motives should be checked against Bloom’s notes and contemporaneous documents where consequential. ## 44.6 Digital development, power, and sovereignty ### United Nations Conference on Trade and Development. *Digital Economy Report 2019: Value Creation and Capture—Implications for Developing Countries*. https://unctad.org/system/files/official-document/der2019_en.pdf Separates digital participation from the geography of profits, firms, data, and productive capability. ### World Bank. *Digital Progress and Trends Report 2025: Strengthening AI Foundations*. https://www.worldbank.org/en/publication/dptr2025-ai-foundations/report Organizes inclusive AI participation around connectivity, compute, context, and competency. Useful operational complement to diffusion theory. ### “An Intellectual History of Digital Colonialism.” *Journal of Communication* 75, no. 5 (2025): 385–397. https://academic.oup.com/joc/article/75/5/385/8078024 Traces the development and uses of the digital-colonialism concept. Useful for situating rather than automatically endorsing the analogy. ### “Digital Sovereignty and Artificial Intelligence: A Normative Approach.” *Ethics and Information Technology* (2024/2025). https://doi.org/10.1007/s10676-024-09810-5 Argues that sovereignty should concern legitimate authority, not mere control. Particularly useful for distinguishing corporate quasi-sovereignty, state power, consumer benefit, and democratic legitimacy. ### Brookings Institution. *Is AI Sovereignty Possible? Balancing Autonomy and Interdependence* (2026). https://www.brookings.edu/wp-content/uploads/2026/02/20260217_AI_sovereignty_final.pdf Maps dependencies across the AI stack and argues for managed interdependence rather than complete domestic self-sufficiency. Recommends selective investment, supplier diversification, interoperability, open models, and independent evaluation. ## 44.7 SLMs, routing, and deployment ### Lu, Zhenyan, Xiang Li, Dongqi Cai, and coauthors. “Demystifying Small Language Models for Edge Deployment.” *Proceedings of ACL 2025*: 14747–14764. https://aclanthology.org/2025.acl-long.718/ Benchmarks 68 publicly accessible models from 100 million to 5 billion parameters across capability and on-device performance. Finds rapid progress and selected parity with older 7B–8B models, while documenting limited in-context learning and optimization constraints. ### Chen, Lingjiao, Matei Zaharia, and James Zou. “FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance.” 2023. https://arxiv.org/abs/2305.05176 Demonstrates model cascades that selectively use different APIs to reduce cost in experimental settings. Supports hybrid-routing feasibility, not automatic reliability in high-stakes applications. ### Ong, Isaac, Amjad Almahairi, Vincent Wu, and coauthors. “RouteLLM: Learning to Route LLMs with Preference Data.” 2024; revised 2025. https://arxiv.org/abs/2406.18665 Develops learned routing between stronger and weaker models, showing meaningful cost reductions at comparable benchmark quality. ### Stripe. “Stripe agrees to acquire OpenRouter to help businesses optimize token routing and usage.” 2026. https://stripe.com/newsroom/news/stripe-agrees-to-acquire-openrouter Primary corporate announcement showing commercial commitment to multi-provider model routing, metering, and AI-business infrastructure. Corporate strategy statement, not independent evidence about eventual market structure. ## 44.8 Labor-market evidence ### Stanford Digital Economy Lab. “Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence.” 2025, with subsequent revision. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/ Finds an employment gap for young workers in highly AI-exposed occupations, concentrated in more substitution-oriented uses, while emphasizing that results are descriptive early evidence rather than causal proof. Does not directly establish cross-border service displacement. ## 44.9 Distillation and legal enclosure ### Fenwick & West. “DeepSeek, Model Distillation, and the Future of AI IP Protection.” 2025. https://www.fenwick.com/insights/publications/deepseek-model-distillation-and-the-future-of-ai-ip-protection Practitioner analysis of authorized and unauthorized distillation, contract, copyright, computer-access law, and patent strategies. Useful issue map, not settled legal authority. The legal treatment of distillation remains jurisdiction-specific and unresolved. --- # 45. Provenance, scope, and maintenance ## 45.1 Origin This companion emerged from a discussion prompted by John Bloom’s *Eccentric Orbits*. The discussion connected Iridium’s universal-service rhetoric and international telecom political economy to Meta’s geographic monetization structure, WhatsApp’s quasi-public-utility role, Bill Gates’s writing on equitable AI, edge SLMs, model routing, and Jeffrey Ding’s account of technological diffusion. ## 45.2 Purpose The public essay is intentionally concise. This companion preserves: - qualifications omitted for legibility; - the prior-art map; - alternative explanations; - technical and institutional branching paths; - counterarguments; - claims requiring future evidence; - sources for further inquiry. It is not a systematic review, legal opinion, investment recommendation, or settled prediction. ## 45.3 Source hierarchy Prefer sources in this order: 1. primary legislation, regulation, financial disclosure, technical documentation, or institutional record; 2. peer-reviewed academic research; 3. high-quality working papers from identifiable researchers; 4. books and credible third-party analysis; 5. corporate announcements for claims about corporate action or stated strategy; 6. journalism for contemporaneous facts unavailable elsewhere; 7. commentary only for framing or hypotheses. Corporate and advocacy sources should not independently establish claims about social impact. ## 45.4 Temporal sensitivity The following require frequent review: - model and hardware capability; - inference prices; - licensing terms; - provider market share; - laws governing model outputs and distillation; - international AI policy; - employment effects; - platform acquisitions and commercial strategy. Historical and conceptual sections are less time-sensitive. ## 45.5 Known evidence gaps - No public country-level Meta cost accounting sufficient to prove an Indian operating loss. - No comprehensive causal incidence analysis linking Western consumer spending to WhatsApp provision. - Limited independent cross-country replication of digital-welfare estimates. - Limited field evidence comparing SLM, frontier-cloud, and human systems in lower-income settings. - Limited longitudinal evidence on local capability accumulation after AI deployment. - Early and incomplete evidence on labor displacement in service-exporting economies. - Unsettled law and licensing around distillation and model-derived training data. ## 45.6 Revision policy Future revisions should: 1. preserve stable claim IDs from Section 40; 2. record substantive changes in a revision log; 3. update confidence levels separately from prose; 4. archive superseded versions; 5. distinguish newly observed evidence from changed interpretation; 6. validate all links and publication metadata; 7. avoid silently converting hypotheses into facts. ## 45.7 Suggested front-matter fields for the assembled file ```yaml --- title: "Who Pays for Intelligence? — Research Companion" description: "Evidence, counterarguments, and research paths on cross-border financing, consumer surplus, AI sovereignty, and technology diffusion." status: "public research companion" artifact_type: "structured research companion" audience: - "AI assistants" - "researchers" - "readers seeking deeper context" canonical_essay: "/who-pays-for-intelligence" version: "1.0" last_reviewed: "2026-09" license: "All rights reserved unless otherwise stated" --- ``` --- # 46. Final synthesis The central argument survives its strongest qualifications in a narrower and more defensible form. It is not publicly established that American consumers literally subsidize each Indian WhatsApp user on an avoidable-cost basis. What is established is that global platforms can separate the geography of revenue from the geography of benefit. Revenue concentrated in affluent markets finances shared technical systems that create large consumer value elsewhere. Low-revenue users are not passive recipients: they contribute attention, network completeness, strategic position, data, and future commercial possibility. The result is neither charity nor conventional fiscal redistribution. Consumer welfare nonetheless tells only part of the story. A population can receive an indispensable service while possessing little control over its rules, continuity, evolution, or replacement. It can gain consumption value while capturing few profits, taxes, technical skills, firms, or standards-setting power. Consumer surplus, local value capture, and legitimate sovereignty are distinct variables. AI could reproduce the cloud-platform settlement. High-wage enterprise demand may finance cheap global access, while model quality, data, governance, and profits remain concentrated in a few jurisdictions and firms. Universal access could coexist with deeply unequal capability and guarantees. AI may also differ from prior platforms. If useful capabilities can be distilled into small models, combined with local knowledge and current feeds, run on affordable local or regional hardware, legally modified, independently validated, and maintained over time, intelligence can become a transferable productive asset. In that case, rich-market demand would finance not merely a globally consumed service but part of the global distribution of cognitive capital. No single technical achievement secures this outcome. Cheap weights are insufficient without hardware, data, evaluation, institutions, skills, organizational redesign, and access to the physical goods needed to act on advice. Local execution is insufficient if chips, licenses, updates, identity, payments, routing, or escalation remain controlled elsewhere. National ownership is insufficient if power simply moves from a foreign platform to an unaccountable domestic incumbent. Jeffrey Ding’s diffusion framework supplies the decisive final step. The gains from a general-purpose technology accrue not only to those who invent or possess it, but to societies able to adapt it throughout ordinary organizations. AI may help lower its own technical diffusion costs, but it cannot automatically create trustworthy institutions, political authority, capital, supply chains, or legitimate governance. The practical policy distinction is therefore between subsidies that purchase **temporary consumption** and investments that build **transferable, compounding capability**. The latter may include open or replaceable models, local data governance, affordable compute, independent validation, technical education, interoperability, credible exit, and institutions able to redesign work around the technology. The future will probably be hybrid. Routine cognition may run locally; difficult cases may escalate to frontier clouds and human experts. Dependence will not disappear, but it can be made visible, contestable, diversified, and governable. The relevant question is not simply whether intelligence becomes cheap or universally reachable. It is who finances it, who benefits, who captures its productive returns, who can govern and replace it, and which societies develop the capacity to compound it. That is the difference between distributing access to intelligence and distributing power over intelligence.